Search

Self-attention mechanisms and non-local blocks have become crucial building blocks for state-of-the-art neural architectures thanks to their unparalleled ability in capturing long-range dependencies in the input. However their cost is quadratic with the number of spatial positions hence making their use impractical in many real case applications. In this work, we analyze these methods through a polynomial lens, and we show that self-attention can be seen as a special case of a 3 rd order polynomial. Within this polynomial framework, we are able to design polynomial operators capable of accessing the same data pattern of non-local and self-attention blocks while reducing the complexity from quadratic to linear. As a result, we propose two modules (Poly-NL and Poly-SA) that can be used as “drop-in” replacements for more-complex non-local and self-attention layers in state-of-the-art CNNs and ViT architectures. Our modules can achieve comparable, if not better, performance across a wide range of computer vision tasks while keeping a complexity equivalent to a standard linear layer.

We introduce a new image segmentation task, called Entity Segmentation (ES), which aims to segment all visual entities (objects and stuffs) in an image without predicting their semantic labels. By removing the need of class label prediction, the models trained for such task can focus more on improving segmentation quality. It has many practical applications such as image manipulation and editing where the quality of segmentation masks is crucial but class labels are less important. We conduct the first-ever study to investigate the feasibility of convolutional center-based representation to segment things and stuffs in a unified manner, and show that such representation fits exceptionally well in the context of ES. More specifically, we propose a CondInst-like fully-convolutional architecture with two novel modules specifically designed to exploit the class-agnostic and non-overlapping requirements of ES. Experiments show that the models designed and trained for ES significantly outperforms popular class-specific panoptic segmentation models in terms of segmentation quality. Moreover, an ES model can be easily trained on a combination of multiple datasets without the need to resolve label conflicts in dataset merging, and the model trained for ES on one or more datasets can generalize very well to other test datasets of unseen domains. The code has been released at https://github.com/dvlab-research/Entity

In computer vision, some attribution methods for explaining CNNs attempt to study how the intermediate features affect network prediction. However, they usually ignore the feature hierarchies among the intermediate features. This paper introduces a hierarchical decomposition framework to explain CNN’s decision-making process in a top-down manner. Specifically, we propose a gradient-based activation propagation (gAP) module that can decompose any intermediate CNN decision to its lower layers and find the supporting features. Then we utilize the gAP module to iteratively decompose the network decision to the supporting evidence from different CNN layers. The proposed framework can generate a deep hierarchy of strongly associated supporting evidence for the network decision, which provides insight into the decision-making process. Moreover, gAP is effort-free for understanding CNN-based models without network architecture modification and extra training processes. Experiments show the effectiveness of the proposed method. The data and source code will be publicly available at https://mmcheng.net/hdecomp/.

Transformers with powerful global relation modeling abilities have been introduced to fundamental computer vision tasks recently. As a typical example, the Vision Transformer (ViT) directly applies a pure transformer architecture on image classification, by simply splitting images into tokens with a fixed length, and employing transformers to learn relations between these tokens. However, such naive tokenization could destruct object structures, assign grids to uninterested regions such as background, and introduce interference signals. To mitigate the above issues, in this paper, we propose an iterative and progressive sampling strategy to locate discriminative regions. At each iteration, embeddings of the current sampling step are fed into a transformer encoder layer, and a group of sampling offsets is predicted to update the sampling locations for the next step. The progressive sampling is differentiable. When combined with the Vision Transformer, the obtained PS-ViT network can adaptively learn where to look. The proposed PS-ViT is both effective and efficient. When trained from scratch on ImageNet, PS-ViT performs 3.8% higher than the vanilla ViT in terms of top-1 accuracy with about $4\times$ fewer parameters and $10\times$ fewer FLOPs. Code is available at https://github.com/yuexy/PS-ViT., Comment: Accepted to ICCV 2021

Image registration is useful for quantifying morphological changes in longitudinal MR images from prostate cancer patients. This paper describes a development in improving the learning-based registration algorithms, for this challenging clinical application often with highly variable yet limited training data. First, we report that the latent space can be clustered into a much lower dimensional space than that commonly found as bottleneck features at the deep layer of a trained registration network. Based on this observation, we propose a hierarchical quantization method, discretizing the learned feature vectors using a jointly-trained dictionary with a constrained size, in order to improve the generalisation of the registration networks. Furthermore, a novel collaborative dictionary is independently optimised to incorporate additional prior information, such as the segmentation of the gland or other regions of interest, in the latent quantized space. Based on 216 real clinical images from 86 prostate cancer patients, we show the efficacy of both the designed components. Improved registration accuracy was obtained with statistical significance, in terms of both Dice on gland and target registration error on corresponding landmarks, the latter of which achieved 5.46 mm, an improvement of 28.7\% from the baseline without quantization. Experimental results also show that the difference in performance was indeed minimised between training and testing data., Comment: preprint version, accepted for MICCAI 2022 (25th International Conference on Medical Image Computing and Computer Assisted Intervention)

Remote photoplethysmography (rPPG), which aims at measuring heart activities and physiological signals from facial video without any contact, has great potential in many applications (e.g., remote healthcare and affective computing). Recent deep learning approaches focus on mining subtle rPPG clues using convolutional neural networks with limited spatio-temporal receptive fields, which neglect the long-range spatio-temporal perception and interaction for rPPG modeling. In this paper, we propose the PhysFormer, an end-to-end video transformer based architecture, to adaptively aggregate both local and global spatio-temporal features for rPPG representation enhancement. As key modules in PhysFormer, the temporal difference transformers first enhance the quasi-periodic rPPG features with temporal difference guided global attention, and then refine the local spatio-temporal representation against interference. Furthermore, we also propose the label distribution learning and a curriculum learning inspired dynamic constraint in frequency domain, which provide elaborate supervisions for PhysFormer and alleviate overfitting. Comprehensive experiments are performed on four benchmark datasets to show our superior performance on both intra- and cross-dataset testings. One highlight is that, unlike most transformer networks needed pretraining from large-scale datasets, the proposed PhysFormer can be easily trained from scratch on rPPG datasets, which makes it promising as a novel transformer baseline for the rPPG community. The codes will be released at https://github.com/ZitongYu/PhysFormer., Accepted by CVPR2022

Empowered by large datasets, e.g., ImageNet, unsupervised learning on large-scale data has enabled significant advances for classification tasks. However, whether the large-scale unsupervised semantic segmentation can be achieved remains unknown. There are two major challenges: i) we need a large-scale benchmark for assessing algorithms; ii) we need to develop methods to simultaneously learn category and shape representation in an unsupervised manner. In this work, we propose a new problem of large-scale unsupervised semantic segmentation (LUSS) with a newly created benchmark dataset to help the research progress. Building on the ImageNet dataset, we propose the ImageNet-S dataset with 1.2 million training images and 50k high-quality semantic segmentation annotations for evaluation. Our benchmark has a high data diversity and a clear task objective. We also present a simple yet effective method that works surprisingly well for LUSS. In addition, we benchmark related un/weakly/fully supervised methods accordingly, identifying the challenges and possible directions of LUSS. The benchmark and source code is publicly available at https://github.com/LUSSeg., Benchmark and Source Code: https://github.com/LUSSeg

Mixup-based augmentation has been found to be effective for generalizing models during training, especially for Vision Transformers (ViTs) since they can easily overfit. However, previous mixup-based methods have an underlying prior knowledge that the linearly interpolated ratio of targets should be kept the same as the ratio proposed in input interpolation. This may lead to a strange phenomenon that sometimes there is no valid object in the mixed image due to the random process in augmentation but there is still response in the label space. To bridge such gap between the input and label spaces, we propose TransMix, which mixes labels based on the attention maps of Vision Transformers. The confidence of the label will be larger if the corresponding input image is weighted higher by the attention map. TransMix is embarrassingly simple and can be implemented in just a few lines of code without introducing any extra parameters and FLOPs to ViT-based models. Experimental results show that our method can consistently improve various ViT-based models at scales on ImageNet classification. After pre-trained with TransMix on ImageNet, the ViT-based models also demonstrate better transferability to semantic segmentation, object detection and instance segmentation. TransMix also exhibits to be more robust when evaluating on 4 different benchmarks. Code will be made publicly available at https://github.com/Beckschen/TransMix., Comment: Code will be made publicly available at https://github.com/Beckschen/TransMix

Recurrent models are a popular choice for video enhancement tasks such as video denoising or super-resolution. In this work, we focus on their stability as dynamical systems and show that they tend to fail catastrophically at inference time on long video sequences. To address this issue, we (1) introduce a diagnostic tool which produces input sequences optimized to trigger instabilities and that can be interpreted as visualizations of temporal receptive fields, and (2) propose two approaches to enforce the stability of a model during training: constraining the spectral norm or constraining the stable rank of its convolutional layers. We then introduce Stable Rank Normalization for Convolutional layers (SRN-C), a new algorithm that enforces these constraints. Our experimental results suggest that SRN-C successfully enforces stablility in recurrent video processing models without a significant performance loss.

International audience; In this paper, we address the problem of surface tracking in multiple camera environments and over time sequences. In order to fully track a surface undergoing significant deformations, we cast the problem as a mesh evolution over time. Such an evolution is driven by 3D displacement fields estimated between meshes recovered independently at different time frames. Geometric and photometric information is used to identify a robust set of matching vertices. This provides a sparse displacement field that is densified over the mesh by Laplacian diffusion. In contrast to existing approaches that evolve meshes, we do not assume a known model or a fixed topology. The contribution is a novel mesh evolution based framework that allows to fully track, over long sequences, an unknown surface encountering deformations, including topological changes. Results on very challenging and publicly available image based 3D mesh sequences demonstrate the ability of our framework to efficiently recover surface motions .

International audience; This paper improves recent methods for large scale image search. State-of-the-art methods build on the bag-of-features image representation. We, first, analyze bag-of-features in the framework of approximate nearest neighbor search. This shows the sub-optimality of such a representation for matching descriptors and leads us to derive a more precise representation based on 1) Hamming embedding (HE) and 2) weak geometric consistency constraints (WGC). HE provides binary signatures that refine the matching based on visual words. WGC filters matching descriptors that are not consistent in terms of angle and scale. HE and WGC are integrated within the inverted file and are efficiently exploited for all images, even in the case of very large datasets. Experiments performed on a dataset of one million of images show a significant improvement due to the binary signature and the weak geometric consistency constraints, as well as their efficiency. Estimation of the full geometric transformation, i.e., a re-ranking step on a short list of images, is complementary to our weak geometric consistency constraints and allows to further improve the accuracy.

International audience; Catadioptric cameras are a popular type of omnidirectional imaging system. Their imaging and multi-view geometry has been exten- sively studied; epipolar geometry for instance, is geometrically speaking, well understood. However, the existence of a bilinear matching constraint and an associated fundamental matrix, has so far only been shown for the special case of para-catadioptric cameras (consisting of a paraboloidal mirror and an orthographic camera). The main goal of this work is to obtain such results for all central catadioptric cameras. Our main result is to show the existence of a general 15x15 fundamental matrix. This is based on and completed by a number of other results, e.g. the formulation of general catadioptric projection matrices and plane homographies.

International audience; In this paper we are interested in finding images of people on the web, and more specifically within large databases of captioned news images. It has recently been shown that visual analysis of the faces in images returned on a text-based query over captions can significantly improve search results. The underlying idea to improve the text-based results is that although this initial result is imperfect, it will render the queried person to be relatively frequent as compared to other people, so we can search for a large group of highly similar faces. The performance of such methods depends strongly on this assumption: for people whose face appears in less than about 40% of the initial text-based result, the performance may be very poor. The contribution of this paper is to improve search results by exploiting faces of other people that co-occur frequently with the queried person. We refer to this process as 'query expansion'. In the face analysis we use the query expansion to provide a query-specific relevant set of 'negative' examples which should be separated from the potentially positive examples in the text-based result set. We apply this idea to a recently-proposed method which filters the initial result set using a Gaussian mixture model, and apply the same idea using a logistic discriminant model. We experimentally evaluate the methods using a set of 23 queries on a database of 15.000 captioned news stories from yahoonews. The results show that (i) query expansion improves both methods, (ii) that our discriminative models outperform the generative ones, and (iii) our best results surpass the state-of-the-art results by 10% precision on average.

International audience; This article proposes a flexible tracker which can estimate motion and deformations of 3D objects by considering their appearances as nonrigid surfaces. In this approach, a flexible model is built by matching features (key points) over training sequences and by learning the deformations of a spline based model. This statistical model captures the variations in the appearance of objects caused by 3D pose variations. Visual tracking is then possible, for each new frame, by matching local features of the model according to their local appearances as well as optimal optimization of the constraints provided by the flexible model. The approach is demonstrated on real-world images sequences.

International audience; This paper describes a new probabilistic framework for recognizing textures in images. Images are described by local affine-invariant descriptors and their spatial relationships. We introduce a statistical parametric models of the dependence between descriptors. We use Hidden Markov Models (HMM) and estimate the parameters with a recent technique based on the mean field principle. Preliminary results for texture recognition are promising and outperform existing techniques.