Search

Domain generalization (DG) methods aim to maintain good performance in an unseen target domain by using training data from multiple source domains. While success on certain occasions are observed, enhancing the baseline across most scenarios remains challenging. This work introduces a simple yet effective framework, dubbed learning from multiple experts (LFME), that aims to make the target model an expert in all source domains to improve DG. Specifically, besides learning the target model used in inference, LFME will also train multiple experts specialized in different domains, whose output probabilities provide professional guidance by simply regularizing the logit of the target model. Delving deep into the framework, we reveal that the introduced logit regularization term implicitly provides effects of enabling the target model to harness more information, and mining hard samples from the experts during training. Extensive experiments on benchmarks from different DG tasks demonstrate that LFME is consistently beneficial to the baseline and can achieve comparable performance to existing arts. Code is available at~\url{https://github.com/liangchen527/LFME}., Comment: Accepted by NeurIPS 2024

Zero-shot subject-driven image generation aims to produce images that incorporate a subject from a given example image. The challenge lies in preserving the subject's identity while aligning with the text prompt which often requires modifying certain aspects of the subject's appearance. Despite advancements in diffusion model based methods, existing approaches still struggle to balance identity preservation with text prompt alignment. In this study, we conducted an in-depth investigation into this issue and uncovered key insights for achieving effective identity preservation while maintaining a strong balance. Our key findings include: (1) the design of the subject image encoder significantly impacts identity preservation quality, and (2) separating text and subject guidance is crucial for both text alignment and identity preservation. Building on these insights, we introduce a new approach called EZIGen, which employs two main strategies: a carefully crafted subject image Encoder based on the pretrained UNet of the Stable Diffusion model to ensure high-quality identity transfer, following a process that decouples the guidance stages and iteratively refines the initial image layout. Through these strategies, EZIGen achieves state-of-the-art results on multiple subject-driven benchmarks with a unified model and 100 times less training data. The demo page is available at: https://zichengduan.github.io/pages/EZIGen/index.html.

Harnessing the robust capabilities of Large Language Models (LLMs) for narrative generation, logical reasoning, and common-sense knowledge integration, this study delves into utilizing LLMs to enhance automated radiology report generation (R2Gen). Despite the wealth of knowledge within LLMs, efficiently triggering relevant knowledge within these large models for specific tasks like R2Gen poses a critical research challenge. This paper presents KARGEN, a Knowledge-enhanced Automated radiology Report GENeration framework based on LLMs. Utilizing a frozen LLM to generate reports, the framework integrates a knowledge graph to unlock chest disease-related knowledge within the LLM to enhance the clinical utility of generated reports. This is achieved by leveraging the knowledge graph to distill disease-related features in a designed way. Since a radiology report encompasses both normal and disease-related findings, the extracted graph-enhanced disease-related features are integrated with regional image features, attending to both aspects. We explore two fusion methods to automatically prioritize and select the most relevant features. The fused features are employed by LLM to generate reports that are more sensitive to diseases and of improved quality. Our approach demonstrates promising results on the MIMIC-CXR and IU-Xray datasets.

Training a fine-grained image recognition model with limited data presents a significant challenge, as the subtle differences between categories may not be easily discernible amidst distracting noise patterns. One commonly employed strategy is to leverage pretrained neural networks, which can generate effective feature representations for constructing an image classification model with a restricted dataset. However, these pretrained neural networks are typically trained for different tasks than the fine-grained visual recognition (FGVR) task at hand, which can lead to the extraction of less relevant features. Moreover, in the context of building FGVR models with limited data, these irrelevant features can dominate the training process, overshadowing more useful, generalizable discriminative features. Our research has identified a surprisingly simple solution to this challenge: we introduce a regularization technique to ensure that the magnitudes of the extracted features are evenly distributed. This regularization is achieved by maximizing the uniformity of feature magnitude distribution, measured through the entropy of the normalized features. The motivation behind this regularization is to remove bias in feature magnitudes from pretrained models, where some features may be more prominent and, consequently, more likely to be used for classification. Additionally, we have developed a dynamic weighting mechanism to adjust the strength of this regularization throughout the learning process. Despite its apparent simplicity, our approach has demonstrated significant performance improvements across various fine-grained visual recognition datasets.

Diffusion models have exhibited remarkable prowess in visual generalization. Building on this success, we introduce an instruction-based object addition pipeline, named Add-SD, which automatically inserts objects into realistic scenes with rational sizes and positions. Different from layout-conditioned methods, Add-SD is solely conditioned on simple text prompts rather than any other human-costly references like bounding boxes. Our work contributes in three aspects: proposing a dataset containing numerous instructed image pairs; fine-tuning a diffusion model for rational generation; and generating synthetic data to boost downstream tasks. The first aspect involves creating a RemovalDataset consisting of original-edited image pairs with textual instructions, where an object has been removed from the original image while maintaining strong pixel consistency in the background. These data pairs are then used for fine-tuning the Stable Diffusion (SD) model. Subsequently, the pretrained Add-SD model allows for the insertion of expected objects into an image with good rationale. Additionally, we generate synthetic instances for downstream task datasets at scale, particularly for tail classes, to alleviate the long-tailed problem. Downstream tasks benefit from the enriched dataset with enhanced diversity and rationale. Experiments on LVIS val demonstrate that Add-SD yields an improvement of 4.3 mAP on rare classes over the baseline. Code and models are available at https://github.com/ylingfeng/Add-SD.

Self-Supervised Learning (SSL) has become a prominent approach for acquiring visual representations across various tasks, yet its application in fine-grained visual recognition (FGVR) is challenged by the intricate task of distinguishing subtle differences between categories. To overcome this, we introduce an novel strategy that boosts SSL's ability to extract critical discriminative features vital for FGVR. This approach creates synthesized data pairs to guide the model to focus on discriminative features critical for FGVR during SSL. We start by identifying non-discriminative features using two main criteria: features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss. We then introduce perturbations to these non-discriminative features while preserving discriminative ones. A decoder is employed to reconstruct images from both perturbed and original feature vectors to create data pairs. An encoder is trained on such generated data pairs to become invariant to variations in non-discriminative dimensions while focusing on discriminative features, thereby improving the model's performance in FGVR tasks. We demonstrate the promising FGVR performance of the proposed approach through extensive evaluation on a wide variety of datasets., Comment: Accepted by ECCV 2024

This paper explores the problem of continual learning (CL) of vision-language models (VLMs) in open domains, where the models need to perform continual updating and inference on a streaming of datasets from diverse seen and unseen domains with novel classes. Such a capability is crucial for various applications in open environments, e.g., AI assistants, autonomous driving systems, and robotics. Current CL studies mostly focus on closed-set scenarios in a single domain with known classes. Large pre-trained VLMs like CLIP have demonstrated superior zero-shot recognition ability, and a number of recent studies leverage this ability to mitigate catastrophic forgetting in CL, but they focus on closed-set CL in a single domain dataset. Open-domain CL of large VLMs is significantly more challenging due to 1) large class correlations and domain gaps across the datasets and 2) the forgetting of zero-shot knowledge in the pre-trained VLMs in addition to the knowledge learned from the newly adapted datasets. In this work we introduce a novel approach, termed CoLeCLIP, that learns an open-domain CL model based on CLIP. It addresses these challenges by a joint learning of a set of task prompts and a cross-domain class vocabulary. Extensive experiments on 11 domain datasets show that CoLeCLIP outperforms state-of-the-art methods for open-domain CL under both task- and class-incremental learning settings.

Learning domain-invariant semantic representations is crucial for achieving domain generalization (DG), where a model is required to perform well on unseen target domains. One critical challenge is that standard training often results in entangled semantic and domain-specific features. Previous works suggest formulating the problem from a causal perspective and solving the entanglement problem by enforcing marginal independence between the causal (\ie semantic) and non-causal (\ie domain-specific) features. Despite its simplicity, the basic marginal independent-based idea alone may be insufficient to identify the causal feature. By d-separation, we observe that the causal feature can be further characterized by being independent of the domain conditioned on the object, and we propose the following two strategies as complements for the basic framework. First, the observation implicitly implies that for the same object, the causal feature should not be associated with the non-causal feature, revealing that the common practice of obtaining the two features with a shared base feature extractor and two lightweight prediction heads might be inappropriate. To meet the constraint, we propose a simple early-branching structure, where the causal and non-causal feature obtaining branches share the first few blocks while diverging thereafter, for better structure design; Second, the observation implies that the causal feature remains invariant across different domains for the same object. To this end, we suggest that augmentation should be incorporated into the framework to better characterize the causal feature, and we further suggest an effective random domain sampling scheme to fulfill the task. Theoretical and experimental results show that the two strategies are beneficial for the basic marginal independent-based framework. Code is available at \url{https://github.com/liangchen527/CausEB}., Comment: Accepted by IJCV

Medical vision language pre-training (VLP) has emerged as a frontier of research, enabling zero-shot pathological recognition by comparing the query image with the textual descriptions for each disease. Due to the complex semantics of biomedical texts, current methods struggle to align medical images with key pathological findings in unstructured reports. This leads to the misalignment with the target disease's textual representation. In this paper, we introduce a novel VLP framework designed to dissect disease descriptions into their fundamental aspects, leveraging prior knowledge about the visual manifestations of pathologies. This is achieved by consulting a large language model and medical experts. Integrating a Transformer module, our approach aligns an input image with the diverse elements of a disease, generating aspect-centric image representations. By consolidating the matches from each aspect, we improve the compatibility between an image and its associated disease. Additionally, capitalizing on the aspect-oriented representations, we present a dual-head Transformer tailored to process known and unknown diseases, optimizing the comprehensive detection efficacy. Conducting experiments on seven downstream datasets, ours improves the accuracy of recent methods by up to 8.56% and 17.26% for seen and unseen categories, respectively. Our code is released at https://github.com/HieuPhan33/MAVL., Comment: Accepted at CVPR2024. Pre-print before final camera-ready version

Class-Agnostic Counting (CAC) seeks to accurately count objects in a given image with only a few reference examples. While previous methods achieving this relied on additional training, recent efforts have shown that it's possible to accomplish this without training by utilizing pre-existing foundation models, particularly the Segment Anything Model (SAM), for counting via instance-level segmentation. Although promising, current training-free methods still lag behind their training-based counterparts in terms of performance. In this research, we present a straightforward training-free solution that effectively bridges this performance gap, serving as a strong baseline. The primary contribution of our work lies in the discovery of four key technologies that can enhance performance. Specifically, we suggest employing a superpixel algorithm to generate more precise initial point prompts, utilizing an image encoder with richer semantic knowledge to replace the SAM encoder for representing candidate objects, and adopting a multiscale mechanism and a transductive prototype scheme to update the representation of reference examples. By combining these four technologies, our approach achieves significant improvements over existing training-free methods and delivers performance on par with training-based ones.

Source-Free Unsupervised Domain Adaptation (SFUDA) is a challenging task where a model needs to be adapted to a new domain without access to target domain labels or source domain data. The primary difficulty in this task is that the model's predictions may be inaccurate, and using these inaccurate predictions for model adaptation can lead to misleading results. To address this issue, this paper proposes a novel approach that considers multiple prediction hypotheses for each sample and investigates the rationale behind each hypothesis. By consolidating these hypothesis rationales, we identify the most likely correct hypotheses, which we then use as a pseudo-labeled set to support a semi-supervised learning procedure for model adaptation. To achieve the optimal performance, we propose a three-step adaptation process: model pre-adaptation, hypothesis consolidation, and semi-supervised learning. Extensive experimental results demonstrate that our approach achieves state-of-the-art performance in the SFUDA task and can be easily integrated into existing approaches to improve their performance. The codes are available at \url{https://github.com/GANPerf/HCPR}.

This work conducts an evaluation of GPT-4V's multimodal capability for medical image analysis, with a focus on three representative tasks of radiology report generation, medical visual question answering, and medical visual grounding. For the evaluation, a set of prompts is designed for each task to induce the corresponding capability of GPT-4V to produce sufficiently good outputs. Three evaluation ways including quantitative analysis, human evaluation, and case study are employed to achieve an in-depth and extensive evaluation. Our evaluation shows that GPT-4V excels in understanding medical images and is able to generate high-quality radiology reports and effectively answer questions about medical images. Meanwhile, it is found that its performance for medical visual grounding needs to be substantially improved. In addition, we observe the discrepancy between the evaluation outcome from quantitative analysis and that from human evaluation. This discrepancy suggests the limitations of conventional metrics in assessing the performance of large language models like GPT-4V and the necessity of developing new metrics for automatic quantitative analysis.

Effective object detection in autonomous vehicles is challenged by deployment in diverse and unfamiliar environments. Online Source-Free Domain Adaptation (O-SFDA) offers model adaptation using a stream of unlabeled data from a target domain in an online manner. However, not all captured frames contain information beneficial for adaptation, especially in the presence of redundant data and class imbalance issues. This paper introduces a novel approach to enhance O-SFDA for adaptive object detection through unsupervised data acquisition. Our methodology prioritizes the most informative unlabeled frames for inclusion in the online training process. Empirical evaluation on a real-world dataset reveals that our method outperforms existing state-of-the-art O-SFDA techniques, demonstrating the viability of unsupervised data acquisition for improving the adaptive object detector., Comment: Accepted by ECCV workshop ROAM 2024; 12 pages, 2 figures

Large Language Models (LLMs) have consistently showcased remarkable generalization capabilities when applied to various language tasks. Nonetheless, harnessing the full potential of LLMs for Radiology Report Generation (R2Gen) still presents a challenge, stemming from the inherent disparity in modality between LLMs and the R2Gen task. To bridge this gap effectively, we propose R2GenGPT, which is a novel solution that aligns visual features with the word embedding space of LLMs using an efficient visual alignment module. This innovative approach empowers the previously static LLM to seamlessly integrate and process image information, marking a step forward in optimizing R2Gen performance. R2GenGPT offers the following benefits. First, it attains state-of-the-art (SOTA) performance by training only the lightweight visual alignment module while freezing all the parameters of LLM. Second, it exhibits high training efficiency, as it requires the training of an exceptionally minimal number of parameters while achieving rapid convergence. By employing delta tuning, our model only trains 5M parameters (which constitute just 0.07\% of the total parameter count) to achieve performance close to the SOTA levels. Our code is available at https://github.com/wang-zhanyu/R2GenGPT., Comment: Accepted by meta-radiology

As an effective way to alleviate the burden of data annotation, semi-supervised learning (SSL) provides an attractive solution due to its ability to leverage both labeled and unlabeled data to build a predictive model. While significant progress has been made recently, SSL algorithms are often evaluated and developed under the assumption that the network is randomly initialized. This is in sharp contrast to most vision recognition systems that are built from fine-tuning a pretrained network for better performance. While the marriage of SSL and a pretrained model seems to be straightforward, recent literature suggests that naively applying state-of-the-art SSL with a pretrained model fails to unleash the full potential of training data. In this paper, we postulate the underlying reason is that the pretrained feature representation could bring a bias inherited from the source data, and the bias tends to be magnified through the self-training process in a typical SSL algorithm. To overcome this issue, we propose to use pseudo-labels from the unlabelled data to update the feature extractor that is less sensitive to incorrect labels and only allow the classifier to be trained from the labeled data. More specifically, we progressively adjust the feature extractor to ensure its induced feature distribution maintains a good class separability even under strong input perturbation. Through extensive experimental studies, we show that the proposed approach achieves superior performance over existing solutions., Comment: to appear at ICCVW2023 (Workshop on Visual Continual Learning)

This paper offers a new perspective to ease the challenge of domain generalization, which involves maintaining robust results even in unseen environments. Our design focuses on the decision-making process in the final classifier layer. Specifically, we propose treating the element-wise contributions to the final results as the rationale for making a decision and representing the rationale for each sample as a matrix. For a well-generalized model, we suggest the rationale matrices for samples belonging to the same category should be similar, indicating the model relies on domain-invariant clues to make decisions, thereby ensuring robust results. To implement this idea, we introduce a rationale invariance loss as a simple regularization technique, requiring only a few lines of code. Our experiments demonstrate that the proposed approach achieves competitive results across various datasets, despite its simplicity. Code is available at \url{https://github.com/liangchen527/RIDG}., Comment: Accepted in ICCV 2023

Compositional Zero-Shot Learning (CZSL) aims to train models to recognize novel compositional concepts based on learned concepts such as attribute-object combinations. One of the challenges is to model attributes interacted with different objects, e.g., the attribute ``wet" in ``wet apple" and ``wet cat" is different. As a solution, we provide analysis and argue that attributes are conditioned on the recognized object and input image and explore learning conditional attribute embeddings by a proposed attribute learning framework containing an attribute hyper learner and an attribute base learner. By encoding conditional attributes, our model enables to generate flexible attribute embeddings for generalization from seen to unseen compositions. Experiments on CZSL benchmarks, including the more challenging C-GQA dataset, demonstrate better performances compared with other state-of-the-art approaches and validate the importance of learning conditional attributes. Code is available at https://github.com/wqshmzh/CANet-CZSL, Comment: 10 pages, 4 figures, accepted in CVPR2023

The Stable Diffusion model is a prominent text-to-image generation model that relies on a text prompt as its input, which is encoded using the Contrastive Language-Image Pre-Training (CLIP). However, text prompts have limitations when it comes to incorporating implicit information from reference images. Existing methods have attempted to address this limitation by employing expensive training procedures involving millions of training samples for image-to-image generation. In contrast, this paper demonstrates that the CLIP model, as utilized in Stable Diffusion, inherently possesses the ability to instantaneously convert images into text prompts. Such an image-to-prompt conversion can be achieved by utilizing a linear projection matrix that is calculated in a closed form. Moreover, the paper showcases that this capability can be further enhanced by either utilizing a small amount of similar-domain training data (approximately 100 images) or incorporating several online training steps (around 30 iterations) on the reference images. By leveraging these approaches, the proposed method offers a simple and flexible solution to bridge the gap between images and text prompts. This methodology can be applied to various tasks such as image variation and image editing, facilitating more effective and seamless interaction between images and textual prompts., Comment: Accepted by NeurIPS 2023, 21 pages, 28 figures

The main challenge in domain generalization (DG) is to handle the distribution shift problem that lies between the training and test data. Recent studies suggest that test-time training (TTT), which adapts the learned model with test data, might be a promising solution to the problem. Generally, a TTT strategy hinges its performance on two main factors: selecting an appropriate auxiliary TTT task for updating and identifying reliable parameters to update during the test phase. Both previous arts and our experiments indicate that TTT may not improve but be detrimental to the learned model if those two factors are not properly considered. This work addresses those two factors by proposing an Improved Test-Time Adaptation (ITTA) method. First, instead of heuristically defining an auxiliary objective, we propose a learnable consistency loss for the TTT task, which contains learnable parameters that can be adjusted toward better alignment between our TTT task and the main prediction task. Second, we introduce additional adaptive parameters for the trained model, and we suggest only updating the adaptive parameters during the test phase. Through extensive experiments, we show that the proposed two strategies are beneficial for the learned model (see Figure 1), and ITTA could achieve superior performance to the current state-of-the-art methods on several DG benchmarks. Code is available at https://github.com/liangchen527/ITTA., Comment: Accepted by CVPR 2023

In clinical scenarios, multi-specialist consultation could significantly benefit the diagnosis, especially for intricate cases. This inspires us to explore a "multi-expert joint diagnosis" mechanism to upgrade the existing "single expert" framework commonly seen in the current literature. To this end, we propose METransformer, a method to realize this idea with a transformer-based backbone. The key design of our method is the introduction of multiple learnable "expert" tokens into both the transformer encoder and decoder. In the encoder, each expert token interacts with both vision tokens and other expert tokens to learn to attend different image regions for image representation. These expert tokens are encouraged to capture complementary information by an orthogonal loss that minimizes their overlap. In the decoder, each attended expert token guides the cross-attention between input words and visual tokens, thus influencing the generated report. A metrics-based expert voting strategy is further developed to generate the final report. By the multi-experts concept, our model enjoys the merits of an ensemble-based approach but through a manner that is computationally more efficient and supports more sophisticated interactions among experts. Experimental results demonstrate the promising performance of our proposed model on two widely used benchmarks. Last but not least, the framework-level innovation makes our work ready to incorporate advances on existing "single-expert" models to further improve its performance., Comment: Accepted by CVPR2023

Autoencoding, which aims to reconstruct the input images through a bottleneck latent representation, is one of the classic feature representation learning strategies. It has been shown effective as an auxiliary task for semi-supervised learning but has become less popular as more sophisticated methods have been proposed in recent years. In this paper, we revisit the idea of using image reconstruction as the auxiliary task and incorporate it with a modern semi-supervised semantic segmentation framework. Surprisingly, we discover that such an old idea in semi-supervised learning can produce results competitive with state-of-the-art semantic segmentation algorithms. By visualizing the intermediate layer activations of the image reconstruction module, we show that the feature map channel could correlate well with the semantic concept, which explains why joint training with the reconstruction task is helpful for the segmentation task. Motivated by our observation, we further proposed a modification to the image reconstruction task, aiming to further disentangle the object clue from the background patterns. From experiment evaluation on various datasets, we show that using reconstruction as auxiliary loss can lead to consistent improvements in various datasets and methods. The proposed method can further lead to significant improvement in object-centric segmentation tasks.

Self-supervised learning (SSL) strategies have demonstrated remarkable performance in various recognition tasks. However, both our preliminary investigation and recent studies suggest that they may be less effective in learning representations for fine-grained visual recognition (FGVR) since many features helpful for optimizing SSL objectives are not suitable for characterizing the subtle differences in FGVR. To overcome this issue, we propose learning an additional screening mechanism to identify discriminative clues commonly seen across instances and classes, dubbed as common rationales in this paper. Intuitively, common rationales tend to correspond to the discriminative patterns from the key parts of foreground objects. We show that a common rationale detector can be learned by simply exploiting the GradCAM induced from the SSL objective without using any pre-trained object parts or saliency detectors, making it seamlessly to be integrated with the existing SSL process. Specifically, we fit the GradCAM with a branch with limited fitting capacity, which allows the branch to capture the common rationales and discard the less common discriminative patterns. At the test stage, the branch generates a set of spatial weights to selectively aggregate features representing an instance. Extensive experimental results on four visual tasks demonstrate that the proposed method can lead to a significant improvement in different evaluation settings., Comment: CVPR 2023

Recently, CLIP has been applied to pixel-level zero-shot learning tasks via a two-stage scheme. The general idea is to first generate class-agnostic region proposals and then feed the cropped proposal regions to CLIP to utilize its image-level zero-shot classification capability. While effective, such a scheme requires two image encoders, one for proposal generation and one for CLIP, leading to a complicated pipeline and high computational cost. In this work, we pursue a simpler-and-efficient one-stage solution that directly extends CLIP's zero-shot prediction capability from image to pixel level. Our investigation starts with a straightforward extension as our baseline that generates semantic masks by comparing the similarity between text and patch embeddings extracted from CLIP. However, such a paradigm could heavily overfit the seen classes and fail to generalize to unseen classes. To handle this issue, we propose three simple-but-effective designs and figure out that they can significantly retain the inherent zero-shot capacity of CLIP and improve pixel-level generalization ability. Incorporating those modifications leads to an efficient zero-shot semantic segmentation system called ZegCLIP. Through extensive experiments on three public benchmarks, ZegCLIP demonstrates superior performance, outperforming the state-of-the-art methods by a large margin under both "inductive" and "transductive" zero-shot settings. In addition, compared with the two-stage method, our one-stage ZegCLIP achieves a speedup of about 5 times faster during inference. We release the code at https://github.com/ZiqinZhou66/ZegCLIP.git., Comment: 12 pages, 8 figures

In the current person Re-identification (ReID) methods, most domain generalization works focus on dealing with style differences between domains while largely ignoring unpredictable camera view change, which we identify as another major factor leading to a poor generalization of ReID methods. To tackle the viewpoint change, this work proposes to use a 3D dense pose estimation model and a texture mapping module to map the pedestrian images to canonical view images. Due to the imperfection of the texture mapping module, the canonical view images may lose the discriminative detail clues from the original images, and thus directly using them for ReID will inevitably result in poor performance. To handle this issue, we propose to fuse the original image and canonical view image via a transformer-based module. The key insight of this design is that the cross-attention mechanism in the transformer could be an ideal solution to align the discriminative texture clues from the original image with the canonical view image, which could compensate for the low-quality texture information of the canonical view image. Through extensive experiments, we show that our method can lead to superior performance over the existing approaches in various evaluation settings.

Semi-supervised semantic segmentation requires the model to effectively propagate the label information from limited annotated images to unlabeled ones. A challenge for such a per-pixel prediction task is the large intra-class variation, i.e., regions belonging to the same class may exhibit a very different appearance even in the same picture. This diversity will make the label propagation hard from pixels to pixels. To address this problem, we propose a novel approach to regularize the distribution of within-class features to ease label propagation difficulty. Specifically, our approach encourages the consistency between the prediction from a linear predictor and the output from a prototype-based predictor, which implicitly encourages features from the same pseudo-class to be close to at least one within-class prototype while staying far from the other between-class prototypes. By further incorporating CutMix operations and a carefully-designed prototype maintenance strategy, we create a semi-supervised semantic segmentation algorithm that demonstrates superior performance over the state-of-the-art methods from extensive experimental evaluation on both Pascal VOC and Cityscapes benchmarks., Comment: Accepted to NeurIPS 2022

It is well-known that a deep neural network has a strong fitting capability and can easily achieve a low training error even with randomly assigned class labels. When the number of training samples is small, or the class labels are noisy, networks tend to memorize patterns specific to individual instances to minimize the training error. This leads to the issue of overfitting and poor generalisation performance. This paper explores a remedy by suppressing the network's tendency to rely on instance-specific patterns for empirical error minimisation. The proposed method is based on an adversarial training framework. It suppresses features that can be utilized to identify individual instances among samples within each class. This leads to classifiers only using features that are both discriminative across classes and common within each class. We call our method Adversarial Suppression of Identity Features (ASIF), and demonstrate the usefulness of this technique in boosting generalisation accuracy when faced with small datasets or noisy labels. Our source code is available., Comment: DICTA 2022

Open-World Instance Segmentation (OWIS) is an emerging research topic that aims to segment class-agnostic object instances from images. The mainstream approaches use a two-stage segmentation framework, which first locates the candidate object bounding boxes and then performs instance segmentation. In this work, we instead promote a single-stage framework for OWIS. We argue that the end-to-end training process in the single-stage framework can be more convenient for directly regularizing the localization of class-agnostic object pixels. Based on the single-stage instance segmentation framework, we propose a regularization model to predict foreground pixels and use its relation to instance segmentation to construct a cross-task consistency loss. We show that such a consistency loss could alleviate the problem of incomplete instance annotation -- a common problem in the existing OWIS datasets. We also show that the proposed loss lends itself to an effective solution to semi-supervised OWIS that could be considered an extreme case that all object annotations are absent for some images. Our extensive experiments demonstrate that the proposed method achieves impressive results in both fully-supervised and semi-supervised settings. Compared to SOTA methods, the proposed method significantly improves the $AP_{100}$ score by 4.75\% in UVO$\rightarrow$UVO setting and 4.05\% in COCO$\rightarrow$UVO setting. In the case of semi-supervised learning, our model learned with only 30\% labeled data, even outperforms its fully-supervised counterpart with 50\% labeled data. The code will be released soon.

The challenge of fine-grained visual recognition often lies in discovering the key discriminative regions. While such regions can be automatically identified from a large-scale labeled dataset, a similar method might become less effective when only a few annotations are available. In low data regimes, a network often struggles to choose the correct regions for recognition and tends to overfit spurious correlated patterns from the training data. To tackle this issue, this paper proposes the self-boosting attention mechanism, a novel method for regularizing the network to focus on the key regions shared across samples and classes. Specifically, the proposed method first generates an attention map for each training image, highlighting the discriminative part for identifying the ground-truth object category. Then the generated attention maps are used as pseudo-annotations. The network is enforced to fit them as an auxiliary task. We call this approach the self-boosting attention mechanism (SAM). We also develop a variant by using SAM to create multiple attention maps to pool convolutional maps in a style of bilinear pooling, dubbed SAM-Bilinear. Through extensive experimental studies, we show that both methods can significantly improve fine-grained visual recognition performance on low data regimes and can be incorporated into existing network architectures. The source code is publicly available at: https://github.com/GANPerf/SAM, Comment: To Appear at ECCV 2022

The Contrastive Language-Image Pre-training (CLIP) Model is a recently proposed large-scale pre-train model which attracts increasing attention in the computer vision community. Benefiting from its gigantic image-text training set, the CLIP model has learned outstanding capabilities in zero-shot learning and image-text matching. To boost the recognition performance of CLIP on some target visual concepts, it is often desirable to further update the CLIP model by fine-tuning some classes-of-interest on extra training data. This operation, however, raises an important concern: will the update hurt the zero-shot learning or image-text matching capability of the CLIP, i.e., the catastrophic forgetting issue? If yes, could existing continual learning algorithms be adapted to alleviate the risk of catastrophic forgetting? To answer these questions, this work conducts a systemic study on the continual learning issue of the CLIP model. We construct evaluation protocols to measure the impact of fine-tuning updates and explore different ways to upgrade existing continual learning methods to mitigate the forgetting issue of the CLIP model. Our study reveals the particular challenges of CLIP continual learning problem and lays a foundation for further researches. Moreover, we propose a new algorithm, dubbed Learning without Forgetting via Replayed Vocabulary (VR-LwF), which shows exact effectiveness for alleviating the forgetting issue of the CLIP model., Comment: 12 pages, 5 figures

The cross-resolution person re-identification (CRReID) problem aims to match low-resolution (LR) query identity images against high resolution (HR) gallery images. It is a challenging and practical problem since the query images often suffer from resolution degradation due to the different capturing conditions from real-world cameras. To address this problem, state-of-the-art (SOTA) solutions either learn the resolution-invariant representation or adopt super-resolution (SR) module to recover the missing information from the LR query. This paper explores an alternative SR-free paradigm to directly compare HR and LR images via a dynamic metric, which is adaptive to the resolution of a query image. We realize this idea by learning resolution-adaptive representations for cross-resolution comparison. Specifically, we propose two resolution-adaptive mechanisms. The first one disentangles the resolution-specific information into different sub-vectors in the penultimate layer of the deep neural networks, and thus creates a varying-length representation. To better extract resolution-dependent information, we further propose to learn resolution-adaptive masks for intermediate residual feature blocks. A novel progressive learning strategy is proposed to train those masks properly. These two mechanisms are combined to boost the performance of CRReID. Experimental results show that the proposed method is superior to existing approaches and achieves SOTA performance on multiple CRReID benchmarks., Comment: Under review

Open-set panoptic segmentation (OPS) problem is a new research direction aiming to perform segmentation for both \known classes and \unknown classes, i.e., the objects ("things") that are never annotated in the training set. The main challenges of OPS are twofold: (1) the infinite possibility of the \unknown object appearances makes it difficult to model them from a limited number of training data. (2) at training time, we are only provided with the "void" category, which essentially mixes the "unknown thing" and "background" classes. We empirically find that directly using "void" category to supervise \known class or "background" classifiers without screening will lead to an unsatisfied OPS result. In this paper, we propose a divide-and-conquer scheme to develop a dual decision process for OPS. We show that by properly combining a \known class discriminator with an additional class-agnostic object prediction head, the OPS performance can be significantly improved. Specifically, we first propose to create a classifier with only \known categories and let the "void" class proposals achieve low prediction probability from those categories. Then we distinguish the "unknown things" from the background by using the additional object prediction head. To further boost performance, we introduce "unknown things" pseudo-labels generated from up-to-date models to enrich the training set. Our extensive experimental evaluation shows that our approach significantly improves \unknown class panoptic quality, with more than 30\% relative improvements than the existing best-performed method.

Recent studies in deepfake detection have yielded promising results when the training and testing face forgeries are from the same dataset. However, the problem remains challenging when one tries to generalize the detector to forgeries created by unseen methods in the training dataset. This work addresses the generalizable deepfake detection from a simple principle: a generalizable representation should be sensitive to diverse types of forgeries. Following this principle, we propose to enrich the "diversity" of forgeries by synthesizing augmented forgeries with a pool of forgery configurations and strengthen the "sensitivity" to the forgeries by enforcing the model to predict the forgery configurations. To effectively explore the large forgery augmentation space, we further propose to use the adversarial training strategy to dynamically synthesize the most challenging forgeries to the current model. Through extensive experiments, we show that the proposed strategies are surprisingly effective (see Figure 1), and they could achieve superior performance than the current state-of-the-art methods. Code is available at \url{https://github.com/liangchen527/SLADD}., Comment: Accepted by CVPR 2022 (Oral presentation)

Deep learning-based person Re-IDentification (ReID) often requires a large amount of training data to achieve good performance. Thus it appears that collecting more training data from diverse environments tends to improve the ReID performance. This paper re-examines this common belief and makes a somehow surprising observation: using more samples, i.e., training with samples from multiple datasets, does not necessarily lead to better performance by using the popular ReID models. In some cases, training with more samples may even hurt the performance of the evaluation is carried out in one of those datasets. We postulate that this phenomenon is due to the incapability of the standard network in adapting to diverse environments. To overcome this issue, we propose an approach called Domain-Camera-Sample Dynamic network (DCSD) whose parameters can be adaptive to various factors. Specifically, we consider the internal domain-related factor that can be identified from the input features, and external domain-related factors, such as domain information or camera information. Our discovery is that training with such an adaptive model can better benefit from more training samples. Experimental results show that our DCSD can greatly boost the performance (up to 12.3%) while joint training in multiple datasets.

Semantic segmentation is a crucial image understanding task, where each pixel of image is categorized into a corresponding label. Since the pixel-wise labeling for ground-truth is tedious and labor intensive, in practical applications, many works exploit the synthetic images to train the model for real-word image semantic segmentation, i.e., Synthetic-to-Real Semantic Segmentation (SRSS). However, Deep Convolutional Neural Networks (CNNs) trained on the source synthetic data may not generalize well to the target real-world data. In this work, we propose two simple yet effective texture randomization mechanisms, Global Texture Randomization (GTR) and Local Texture Randomization (LTR), for Domain Generalization based SRSS. GTR is proposed to randomize the texture of source images into diverse unreal texture styles. It aims to alleviate the reliance of the network on texture while promoting the learning of the domain-invariant cues. In addition, we find the texture difference is not always occurred in entire image and may only appear in some local areas. Therefore, we further propose a LTR mechanism to generate diverse local regions for partially stylizing the source images. Finally, we implement a regularization of Consistency between GTR and LTR (CGL) aiming to harmonize the two proposed mechanisms during training. Extensive experiments on five publicly available datasets (i.e., GTA5, SYNTHIA, Cityscapes, BDDS and Mapillary) with various SRSS settings (i.e., GTA5/SYNTHIA to Cityscapes/BDDS/Mapillary) demonstrate that the proposed method is superior to the state-of-the-art methods for domain generalization based SRSS., Comment: 15 pages, 14 figures, accepted by IEEE Transactions on Image Processing (TIP 2021)

Given a query patch from a novel class, one-shot object detection aims to detect all instances of that class in a target image through the semantic similarity comparison. However, due to the extremely limited guidance in the novel class as well as the unseen appearance difference between query and target instances, it is difficult to appropriately exploit their semantic similarity and generalize well. To mitigate this problem, we present a universal Cross-Attention Transformer (CAT) module for accurate and efficient semantic similarity comparison in one-shot object detection. The proposed CAT utilizes transformer mechanism to comprehensively capture bi-directional correspondence between any paired pixels from the query and the target image, which empowers us to sufficiently exploit their semantic characteristics for accurate similarity comparison. In addition, the proposed CAT enables feature dimensionality compression for inference speedup without performance loss. Extensive experiments on COCO, VOC, and FSOD under one-shot settings demonstrate the effectiveness and efficiency of our method, e.g., it surpasses CoAE, a major baseline in this task by 1.0% in AP on COCO and runs nearly 2.5 times faster. Code will be available in the future., Comment: 7 pages, 3 figures

The training loss function that enforces certain training sample distribution patterns plays a critical role in building a re-identification (ReID) system. Besides the basic requirement of discrimination, i.e., the features corresponding to different identities should not be mixed, additional intra-class distribution constraints, such as features from the same identities should be close to their centers, have been adopted to construct losses. Despite the advances of various new loss functions, it is still challenging to strike the balance between the need of reducing the intra-class variation and allowing certain distribution freedom. In this paper, we propose a new loss based on center predictivity, that is, a sample must be positioned in a location of the feature space such that from it we can roughly predict the location of the center of same-class samples. The prediction error is then regarded as a loss called Center Prediction Loss (CPL). We show that, without introducing additional hyper-parameters, this new loss leads to a more flexible intra-class distribution constraint while ensuring the between-class samples are well-separated. Extensive experiments on various real-world ReID datasets show that the proposed loss can achieve superior performance and can also be complementary to existing losses.

Learning cross-view consistent feature representation is the key for accurate vehicle Re-identification (ReID), since the visual appearance of vehicles changes significantly under different viewpoints. To this end, most existing approaches resort to the supervised cross-view learning using extensive extra viewpoints annotations, which however, is difficult to deploy in real applications due to the expensive labelling cost and the continous viewpoint variation that makes it hard to define discrete viewpoint labels. In this study, we present a pluggable Weakly-supervised Cross-View Learning (WCVL) module for vehicle ReID. Through hallucinating the cross-view samples as the hardest positive counterparts in feature domain, we can learn the consistent feature representation via minimizing the cross-view feature distance based on vehicle IDs only without using any viewpoint annotation. More importantly, the proposed method can be seamlessly plugged into most existing vehicle ReID baselines for cross-view learning without re-training the baselines. To demonstrate its efficacy, we plug the proposed method into a bunch of off-the-shelf baselines and obtain significant performance improvement on four public benchmark datasets, i.e., VeRi-776, VehicleID, VRIC and VRAI.

Fashion products typically feature in compositions of a variety of styles at different clothing parts. In order to distinguish images of different fashion products, we need to extract both appearance (i.e., "how to describe") and localization (i.e.,"where to look") information, and their interactions. To this end, we propose a biologically inspired framework for image-based fashion product retrieval, which mimics the hypothesized twostream visual processing system of human brain. The proposed attentional heterogeneous bilinear network (AHBN) consists of two branches: a deep CNN branch to extract fine-grained appearance attributes and a fully convolutional branch to extract landmark localization information. A joint channel-wise attention mechanism is further applied to the extracted heterogeneous features to focus on important channels, followed by a compact bilinear pooling layer to model the interaction of the two streams. Our proposed framework achieves satisfactory performance on three image-based fashion product retrieval benchmarks., Comment: 11 pages, 8 figures

Most existing crowd counting systems rely on the availability of the object location annotation which can be expensive to obtain. To reduce the annotation cost, one attractive solution is to leverage a large number of unlabeled images to build a crowd counting model in semi-supervised fashion. This paper tackles the semi-supervised crowd counting problem from the perspective of feature learning. Our key idea is to leverage the unlabeled images to train a generic feature extractor rather than the entire network of a crowd counter. The rationale of this design is that learning the feature extractor can be more reliable and robust towards the inevitable noisy supervision generated from the unlabeled data. Also, on top of a good feature extractor, it is possible to build a density map regressor with much fewer density map annotations. Specifically, we proposed a novel semi-supervised crowd counting method which is built upon two innovative components: (1) a set of inter-related binary segmentation tasks are derived from the original density map regression task as the surrogate prediction target; (2) the surrogate target predictors are learned from both labeled and unlabeled data by utilizing a proposed self-training scheme which fully exploits the underlying constraints of these binary segmentation tasks. Through experiments, we show that the proposed method is superior over the existing semisupervised crowd counting method and other representative baselines., Comment: To be appeared in ECCV 2020

Most existing crowd counting methods require object location-level annotation, i.e., placing a dot at the center of an object. While being simpler than the bounding-box or pixel-level annotation, obtaining this annotation is still labor-intensive and time-consuming especially for images with highly crowded scenes. On the other hand, weaker annotations that only know the total count of objects can be almost effortless in many practical scenarios. Thus, it is desirable to develop a learning method that can effectively train models from count-level annotations. To this end, this paper studies the problem of weakly-supervised crowd counting which learns a model from only a small amount of location-level annotations (fully-supervised) but a large amount of count-level annotations (weakly-supervised). To perform effective training in this scenario, we observe that the direct solution of regressing the integral of density map to the object count is not sufficient and it is beneficial to introduce stronger regularizations on the predicted density map of weakly-annotated images. We devise a simple-yet-effective training strategy, namely Multiple Auxiliary Tasks Training (MATT), to construct regularizes for restricting the freedom of the generated density maps. Through extensive experiments on existing datasets and a newly proposed dataset, we validate the effectiveness of the proposed weakly-supervised method and demonstrate its superior performance over existing solutions.

Hyperspectral image(HSI) classification has been improved with convolutional neural network(CNN) in very recent years. Being different from the RGB datasets, different HSI datasets are generally captured by various remote sensors and have different spectral configurations. Moreover, each HSI dataset only contains very limited training samples and thus it is prone to overfitting when using deep CNNs. In this paper, we first deliver a 3D asymmetric inception network, AINet, to overcome the overfitting problem. With the emphasis on spectral signatures over spatial contexts of HSI data, AINet can convey and classify the features effectively. In addition, the proposed data fusion transfer learning strategy is beneficial in boosting the classification performance. Extensive experiments show that the proposed approach beat all of the state-of-art methods on several HSI benchmarks, including Pavia University, Indian Pines and Kennedy Space Center(KSC). Code can be found at: https://github.com/UniLauX/AINet., Comment: 8 pages

Self-supervised learning (SlfSL), aiming at learning feature representations through ingeniously designed pretext tasks without human annotation, has achieved compelling progress in the past few years. Very recently, SlfSL has also been identified as a promising solution for semi-supervised learning (SemSL) since it offers a new paradigm to utilize unlabeled data. This work further explores this direction by proposing to couple SlfSL with SemSL. Our insight is that the prediction target in SemSL can be modeled as the latent factor in the predictor for the SlfSL target. Marginalizing over the latent factor naturally derives a new formulation which marries the prediction targets of these two learning processes. By implementing this idea through a simple-but-effective SlfSL approach -- rotation angle prediction, we create a new SemSL approach called Conditional Rotation Angle Estimation (CRAE). Specifically, CRAE is featured by adopting a module which predicts the image rotation angle conditioned on the candidate image class. Through experimental evaluation, we show that CRAE achieves superior performance over the other existing ways of combining SlfSL and SemSL. To further boost CRAE, we propose two extensions to strengthen the coupling between SemSL target and SlfSL target in basic CRAE. We show that this leads to an improved CRAE method which can achieve the state-of-the-art SemSL performance.

Real-world visual data often exhibits a long-tailed distribution, where some ''head'' classes have a large number of samples, yet only a few samples are available for ''tail'' classes. Such imbalanced distribution causes a great challenge for learning a deep neural network, which can be boiled down into a dilemma: on the one hand, we prefer to increase the exposure of tail class samples to avoid the excessive dominance of head classes in the classifier training. On the other hand, oversampling tail classes makes the network prone to over-fitting, since head class samples are often consequently under-represented. To resolve this dilemma, in this paper, we propose a simple-yet-effective auxiliary learning approach. The key idea is to split a network into a classifier part and a feature extractor part, and then employ different training strategies for each part. Specifically, to promote the awareness of tail-classes, a class-balanced sampling scheme is utilised for training both the classifier and the feature extractor. For the feature extractor, we also introduce an auxiliary training task, which is to train a classifier under the regular random sampling scheme. In this way, the feature extractor is jointly trained from both sampling strategies and thus can take advantage of all training data and avoid the over-fitting issue. Apart from this basic auxiliary task, we further explore the benefit of using self-supervised learning as the auxiliary task. Without using any bells and whistles, our model achieves superior performance over the state-of-the-art solutions.

This paper tackles the problem of video object segmentation. We are specifically concerned with the task of segmenting all pixels of a target object in all frames, given the annotation mask in the first frame. Even when such annotation is available this remains a challenging problem because of the changing appearance and shape of the object over time. In this paper, we tackle this task by formulating it as a meta-learning problem, where the base learner grasping the semantic scene understanding for a general type of objects, and the meta learner quickly adapting the appearance of the target object with a few examples. Our proposed meta-learning method uses a closed form optimizer, the so-called "ridge regression", which has been shown to be conducive for fast and better training convergence. Moreover, we propose a mechanism, named "block splitting", to further speed up the training process as well as to reduce the number of learning parameters. In comparison with the-state-of-the art methods, our proposed framework achieves significant boost up in processing speed, while having very competitive performance compared to the best performing methods on the widely used datasets., Comment: 9 pages

We propose methods to train convolutional neural networks (CNNs) with both binarized weights and activations, leading to quantized models that are specifically friendly to mobile devices with limited power capacity and computation resources. Previous works on quantizing CNNs often seek to approximate the floating-point information using a set of discrete values, which we call value approximation, typically assuming the same architecture as the full-precision networks. Here we take a novel "structure approximation" view of quantization -- it is very likely that different architectures designed for low-bit networks may be better for achieving good performance. In particular, we propose a "network decomposition" strategy, termed Group-Net, in which we divide the network into groups. Thus, each full-precision group can be effectively reconstructed by aggregating a set of homogeneous binary branches. In addition, we learn effective connections among groups to improve the representation capability. Moreover, the proposed Group-Net shows strong generalization to other tasks. For instance, we extend Group-Net for accurate semantic segmentation by embedding rich context into the binary structure. Furthermore, for the first time, we apply binary neural networks to object detection. Experiments on both classification, semantic segmentation and object detection tasks demonstrate the superior performance of the proposed methods over various quantized networks in the literature. Our methods outperform the previous best binary neural networks in terms of accuracy and computation efficiency., Comment: Accepted to Int. J. Comp. Vision (IJCV). Extended version of the conference version arXiv:1811.10413

As a milestone for video object segmentation, one-shot video object segmentation (OSVOS) has achieved a large margin compared to the conventional optical-flow based methods regarding to the segmentation accuracy. Its excellent performance mainly benefit from the three-step training mechanism, that are: (1) acquiring object features on the base dataset (i.e. ImageNet), (2) training the parent network on the training set of the target dataset (i.e. DAVIS-2016) to be capable of differentiating the object of interest from the background. (3) online fine-tuning the interested object on the first frame of the target test set to overfit its appearance, then the model can be utilized to segment the same object in the rest frames of that video. In this paper, we argue that for the step (2), OSVOS has the limitation to 'overemphasize' the generic semantic object information while 'dilute' the instance cues of the object(s), which largely block the whole training process. Through adding a common module, video loss, which we formulate with various forms of constraints (including weighted BCE loss, high-dimensional triplet loss, as well as a novel mixed instance-aware video loss), to train the parent network in the step (2), the network is then better prepared for the step (3), i.e. online fine-tuning on the target instance. Through extensive experiments using different network structures as the backbone, we show that the proposed video loss module can improve the segmentation performance significantly, compared to that of OSVOS. Meanwhile, since video loss is a common module, it can be generalized to other fine-tuning based methods and similar vision tasks such as depth estimation and saliency detection.

This paper tackles the problem of training a deep convolutional neural network of both low-bitwidth weights and activations. Optimizing a low-precision network is very challenging due to the non-differentiability of the quantizer, which may result in substantial accuracy loss. To address this, we propose three practical approaches, including (i) progressive quantization; (ii) stochastic precision; and (iii) joint knowledge distillation to improve the network training. First, for progressive quantization, we propose two schemes to progressively find good local minima. Specifically, we propose to first optimize a net with quantized weights and subsequently quantize activations. This is in contrast to the traditional methods which optimize them simultaneously. Furthermore, we propose a second progressive quantization scheme which gradually decreases the bit-width from high-precision to low-precision during training. Second, to alleviate the excessive training burden due to the multi-round training stages, we further propose a one-stage stochastic precision strategy to randomly sample and quantize sub-networks while keeping other parts in full-precision. Finally, we adopt a novel learning scheme to jointly train a full-precision model alongside the low-precision one. By doing so, the full-precision model provides hints to guide the low-precision model training and significantly improves the performance of the low-precision network. Extensive experiments on various datasets (e.g., CIFAR-100, ImageNet) show the effectiveness of the proposed methods., Comment: Accepted to IEEE T. Pattern Analysis and Machine Intelligence (TPAMI). Extended version of arXiv:1711.00205 (CVPR 2018)

One of the primary challenges faced by deep learning is the degree to which current methods exploit superficial statistics and dataset bias, rather than learning to generalise over the specific representations they have experienced. This is a critical concern because generalisation enables robust reasoning over unseen data, whereas leveraging superficial statistics is fragile to even small changes in data distribution. To illuminate the issue and drive progress towards a solution, we propose a test that explicitly evaluates abstract reasoning over visual data. We introduce a large-scale benchmark of visual questions that involve operations fundamental to many high-level vision tasks, such as comparisons of counts and logical operations on complex visual properties. The benchmark directly measures a method's ability to infer high-level relationships and to generalise them over image-based concepts. It includes multiple training/test splits that require controlled levels of generalization. We evaluate a range of deep learning architectures, and find that existing models, including those popular for vision-and-language tasks, are unable to solve seemingly-simple instances. Models using relational networks fare better but leave substantial room for improvement.

Deep autoencoder has been extensively used for anomaly detection. Training on the normal data, the autoencoder is expected to produce higher reconstruction error for the abnormal inputs than the normal ones, which is adopted as a criterion for identifying anomalies. However, this assumption does not always hold in practice. It has been observed that sometimes the autoencoder "generalizes" so well that it can also reconstruct anomalies well, leading to the miss detection of anomalies. To mitigate this drawback for autoencoder based anomaly detector, we propose to augment the autoencoder with a memory module and develop an improved autoencoder called memory-augmented autoencoder, i.e. MemAE. Given an input, MemAE firstly obtains the encoding from the encoder and then uses it as a query to retrieve the most relevant memory items for reconstruction. At the training stage, the memory contents are updated and are encouraged to represent the prototypical elements of the normal data. At the test stage, the learned memory will be fixed, and the reconstruction is obtained from a few selected memory records of the normal data. The reconstruction will thus tend to be close to a normal sample. Thus the reconstructed errors on anomalies will be strengthened for anomaly detection. MemAE is free of assumptions on the data type and thus general to be applied to different tasks. Experiments on various datasets prove the excellent generalization and high effectiveness of the proposed MemAE., Comment: Accepted to appear at ICCV 2019

In this paper, we seek to tackle a challenge in training low-precision networks: the notorious difficulty in propagating gradient through a low-precision network due to the non-differentiable quantization function. We propose a solution by training the low-precision network with a fullprecision auxiliary module. Specifically, during training, we construct a mix-precision network by augmenting the original low-precision network with the full precision auxiliary module. Then the augmented mix-precision network and the low-precision network are jointly optimized. This strategy creates additional full-precision routes to update the parameters of the low-precision model, thus making the gradient back-propagates more easily. At the inference time, we discard the auxiliary module without introducing any computational complexity to the low-precision network. We evaluate the proposed method on image classification and object detection over various quantization approaches and show consistent performance increase. In particular, we achieve near lossless performance to the full-precision model by using a 4-bit detector, which is of great practical value., Comment: Accepted to Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2020)