Search

Recent methods in network pruning have indicated that a dense neural network involves a sparse subnetwork (called a winning ticket), which can achieve similar test accuracy to its dense counterpart with much fewer network parameters. Generally, these methods search for the winning tickets on well-labeled data. Unfortunately, in many real-world applications, the training data are unavoidably contaminated with noisy labels, thereby leading to performance deterioration of these methods. To address the above-mentioned problem, we propose a novel two-stream sample selection network (TS3-Net), which consists of a sparse subnetwork and a dense subnetwork, to effectively identify the winning ticket with noisy labels. The training of TS3-Net contains an iterative procedure that switches between training both subnetworks and pruning the smallest magnitude weights of the sparse subnetwork. In particular, we develop a multistage learning framework including a warm-up stage, a semisupervised alternate learning stage, and a label refinement stage, to progressively train the two subnetworks. In this way, the classification capability of the sparse subnetwork can be gradually improved at a high sparsity level. Extensive experimental results on both synthetic and real-world noisy datasets (including MNIST, CIFAR-10, CIFAR-100, ANIMAL-10N, Clothing1M, and WebVision) demonstrate that our proposed method achieves state-of-the-art performance with very small memory consumption for label noise learning. Code is available at https://github.com/Runqing-forMost/TS3-Net/tree/master.

Existing caricature-visual face recognition methods train the models based on caricature-visual image pairs from the same identities. Unfortunately, in many real-world applications, facial caricatures and visual facial images are usually unpaired in the training set due to the difficulty of collecting facial caricatures drawn by artists. In this paper, we study caricature-visual face recognition under the practical setting that only unpaired facial caricature and visual facial images are available as training samples, and define this setting as unpaired caricature-visual face recognition. To this end, we develop a novel feature decomposition-restoration-decomposition method (FDRD), which mainly consists of a backbone network, an identity-oriented feature decomposition module, and a modality-oriented feature restoration module, to extract modality-irrelevant identity features. To effectively train FDRD in the case of limited facial caricature training samples, we develop a two-stage learning framework. In the first stage, we perform single-modality restoration, enabling the model to have the basic ability of feature decomposition and restoration for each modality. In the second stage, we perform cross-modality recognition by exchanging new modality features between the two modalities, facilitating the model to focus on the decoupling of identity features and modality features. Experimental results demonstrate that our method performs favorably against several state-of-the-art face recognition methods and cross-modality methods. Our code is available at https://github.com/Capricorn-Karma/FDRD.

Video data bring a big challenge to semantic segmentation due to the large volume of data and strong inter-frame redundancy. In this paper, we propose a dual local and global correlation network tailored for efficient video semantic segmentation. It consists of three modules: 1) a local attention based module, which measures correlation and achieves feature aggregation in a local region between key frame and non-key frame; 2) a consistent constraint module, which considers long-range correlation among pixels from a global view for promoting intra-frame semantic consistency of non-key frame; and 3) a key frame decision module, which selects key frames adaptively based on the ability of feature transferring. Extensive experiments on the Cityscapes and Camvid video datasets demonstrate that our proposed method could reduce inference time significantly while maintaining high accuracy. The implementation is available at https://github.com/An01168/DCNVSS.

Metric learning aims to learn a distance metric such that semantically similar instances are pulled together while dissimilar instances are pushed away. Many existing methods consider maximizing or at least constraining a distance margin in the feature space that separates similar and dissimilar pairs of instances to guarantee their generalization ability. In this article, we advocate imposing an adversarial margin in the input space so as to improve the generalization and robustness of metric learning algorithms. We first show that the adversarial margin, defined as the distance between training instances and their closest adversarial examples in the input space, takes account of both the distance margin in the feature space and the correlation between the metric and triplet constraints. Next, to enhance robustness to instance perturbation, we propose to enlarge the adversarial margin through minimizing a derived novel loss function termed the perturbation loss. The proposed loss can be viewed as a data-dependent regularizer and easily plugged into any existing metric learning methods. Finally, we show that the enlarged margin is beneficial to the generalization ability by using the theoretical technique of algorithmic robustness. Experimental results on 16 datasets demonstrate the superiority of the proposed method over existing state-of-the-art methods in both discrimination accuracy and robustness against possible noise.

Few-shot fine-grained image classification has attracted considerable attention in recent years for its realistic setting to imitate how humans conduct recognition tasks. Metric-based few-shot classifiers have achieved high accuracies. However, their metric function usually requires two arguments of vectors, while transforming or reshaping three-dimensional feature maps to vectors can result in loss of spatial information. Image reconstruction is thus involved to retain more appearance details: the test images are reconstructed by different classes and then classified to the one with the smallest reconstruction error. However, discriminative local information, vital to distinguish sub-categories in fine-grained images with high similarities, is not well elaborated when only the base features from a usual embedding module are adopted for reconstruction. Hence, we propose the novel local content-enriched cross-reconstruction network (LCCRN) for few-shot fine-grained classification. In LCCRN, we design two new modules: the local content-enriched module (LCEM) to learn the discriminative local features, and the cross-reconstruction module (CRM) to fully engage the local features with the appearance details obtained from a separate embedding module. The classification score is calculated based on the weighted sum of reconstruction errors of the cross-reconstruction tasks, with weights learnt from the training process. Extensive experiments on four fine-grained datasets showcase the superior classification performance of LCCRN compared with the state-of-the-art few-shot classification methods. Codes are available at: https://github.com/lutsong/LCCRN.

Person attribute recognition (PAR) aims to simultaneously predict multiple attributes of a person. Existing deep learning-based PAR methods have achieved impressive performance. Unfortunately, these methods usually ignore the fact that different attributes have an imbalance in the number of noisy-labeled samples in the PAR training datasets, thus leading to suboptimal performance. To address the above problem of imbalanced noisy-labeled samples, we propose a novel and effective loss called drop loss for PAR. In the drop loss, the attributes are treated differently in an easy-to-hard way. In particular, the noisy-labeled candidates, which are identified according to their gradient norms, are dropped with a higher drop rate for the harder attribute. Such a manner adaptively alleviates the adverse effect of imbalanced noisy-labeled samples on model learning. To illustrate the effectiveness of the proposed loss, we train a simple ResNet-50 model based on the drop loss and term it DropNet. Experimental results on two representative PAR tasks (including facial attribute recognition and pedestrian attribute recognition) demonstrate that the proposed DropNet achieves comparable or better performance in terms of both balanced accuracy and classification accuracy over several state-of-the-art PAR methods.

Transparent and reflective objects are omnipresent in our daily life, but their unique visual and optical characteristics are notoriously challenging even for state-of-the-art deep networks of semantic segmentation. To alleviate this challenge, we construct a new large-scale real-world RGB-D dataset called TROSD, which is more comprehensive than existing datasets for transparent and reflective object segmentation. Our TROSD dataset contains 11,060 RGB-D images with three semantic classes in terms of transparent objects, reflective objects, and others, covering a variety of daily scenes. Together with the dataset, we also introduce a novel network (TROSNet) as a high-standard baseline to assist other researchers to develop and benchmark their algorithms of transparent and reflective object segmentation. Moreover, extensive experiments also clearly show that the proposed TROSD dataset has an excellent capacity to facilitate the development of semantic segmentation algorithms with strong generalizability.

In this paper, we reveal the discriminant capacity of orthogonal data projection onto the generalized difference subspace (GDS), both theoretically and experimentally. In our previous work, we demonstrated that the GDS projection works as a quasi-orthogonalization of class subspaces, which is an effective feature extraction for subspace based classifiers. Here, we further show that GDS projection also works as a discriminant feature extraction through a similar mechanism to the Fisher discriminant analysis (FDA). A direct proof of the connection between GDS projection and FDA is difficult due to the significant difference in their formulations. To circumvent the complication, we first introduce geometrical Fisher discriminant analysis (gFDA) based on a simplified Fisher criterion. It is derived from a heuristic yet practically plausible assumption: the direction of the sample mean vector of a class is largely aligned to the first principal component vector of the class, given that the principal component analysis (PCA) is applied without data centering. gFDA works stably even under few samples, bypassing the small sample size (SSS) problem of FDA. We then prove that gFDA is equivalent to GDS projection with a small correction term. This equivalence ensures GDS projection to inherit the discriminant ability from FDA via gFDA. Furthermore, we discuss two useful extensions of these methods, 1) a nonlinear extension by kernel trick, 2) a combination with CNN features. The equivalence and the effectiveness of the extensions have been verified through extensive experiments on the extended Yale B+, CMU face database, ALOI, ETH80, MNIST, and CIFAR10, mainly focusing on image recognition under small samples.

As a typical non-Gaussian vector variable, a neutral vector variable contains nonnegative elements only, and its $l_{1}$ -norm equals one. In addition, its neutral properties make it significantly different from the commonly studied vector variables (e.g., the Gaussian vector variables). Due to the aforementioned properties, the conventionally applied linear transformation approaches [e.g., principal component analysis (PCA) and independent component analysis (ICA)] are not suitable for neutral vector variables, as PCA cannot transform a neutral vector variable, which is highly negatively correlated, into a set of mutually independent scalar variables and ICA cannot preserve the bounded property after transformation. In recent work, we proposed an efficient nonlinear transformation approach, i.e., the parallel nonlinear transformation (PNT), for decorrelating neutral vector variables. In this article, we extensively compare PNT with PCA and ICA through both theoretical analysis and experimental evaluations. The results of our investigations demonstrate the superiority of PNT for decorrelating the neutral vector variables.

GAN inversion aims to invert a given image back into the latent space of a pretrained GAN model so that the image can be faithfully reconstructed from the inverted code by the generator. As an emerging technique to bridge the real and fake image domains, GAN inversion plays an essential role in enabling pretrained GAN models, such as StyleGAN and BigGAN, for applications of real image editing. Moreover, GAN inversion interprets GAN's latent space and examines how realistic images can be generated. In this paper, we provide a survey of GAN inversion with a focus on its representative algorithms and its applications in image restoration and image manipulation. We further discuss the trends and challenges for future research. A curated list of GAN inversion methods, datasets, and other related information can be found at https://github.com/weihaox/awesome-gan-inversion.

Object detection has made enormous progress and has been widely used in many applications. However, it performs poorly when only limited training data is available for novel classes that the model has never seen before. Most existing approaches solve few-shot detection tasks implicitly without directly modeling the detectors for novel classes. In this article, we propose GenDet, a new meta-learning-based framework that can effectively generate object detectors for novel classes from few shots and, thus, conducts few-shot detection tasks explicitly. The detector generator is trained by numerous few-shot detection tasks sampled from base classes each with sufficient samples, and thus, it is expected to generalize well on novel classes. An adaptive pooling module is further introduced to suppress distracting samples and aggregate the detectors generated from multiple shots. Moreover, we propose to train a reference detector for each base class in the conventional way, with which to guide the training of the detector generator. The reference detectors and the detector generator can be trained simultaneously. Finally, the generated detectors of different classes are encouraged to be orthogonal to each other for better generalization. The proposed approach is extensively evaluated on the ImageNet, VOC, and COCO data sets under various few-shot detection settings, and it achieves new state-of-the-art results.

In deep face recognition, the commonly used softmax loss and its newly proposed variations are not yet sufficiently effective to handle the class imbalance and softmax saturation issues during the training process while extracting discriminative features. In this brief, to address both issues, we propose a class-variant margin (CVM) normalized softmax loss, by introducing a true-class margin and a false-class margin into the cosine space of the angle between the feature vector and the class-weight vector. The true-class margin alleviates the class imbalance problem, and the false-class margin postpones the early individual saturation of softmax. With negligible computational complexity increment during training, the new loss function is easy to implement in the common deep learning frameworks. Comprehensive experiments on the LFW, YTF, and MegaFace protocols demonstrate the effectiveness of the proposed CVM loss function.

Elastic weight consolidation (EWC) has been successfully applied for general incremental learning to overcome the catastrophic forgetting issue. It adaptively constrains each parameter of the new model not to deviate much from its counterpart in the old model during fine-tuning on new class data sets, according to its importance weight for old tasks. However, the previous study demonstrates that it still suffers from catastrophic forgetting when directly used in object detection. In this article, we show EWC is effective for incremental object detection if with critical adaptations. First, we conduct controlled experiments to identify two core issues why EWC fails if trivially applied to incremental detection: 1) the absence of old class annotations in new class images makes EWC misclassify objects of old classes in these images as background and 2) the quadratic regularization loss in EWC easily leads to gradient explosion when balancing old and new classes. Then, based on the abovementioned findings, we propose the corresponding solutions to tackle these issues: 1) utilize pseudobounding box annotations of old classes on new data sets to compensate for the absence of old class annotations and 2) adopt a novel Huber regularization instead of the original quadratic loss to prevent from unstable training. Finally, we propose a general EWC-based incremental object detection framework and implement it under both Fast R-CNN and Faster R-CNN, showing its flexibility and versatility. In terms of either the final performance or the performance drop with respect to the upper bound of joint training on all seen classes, evaluations on the PASCAL VOC and COCO data sets show that our method achieves a new state of the art.

In human–computer interaction, it is important to accurately estimate the hand pose, especially fingertips. However, traditional approaches to fingertip localization mainly rely on depth images and thus suffer considerably from noise and missing values. Instead of depth images, stereo images can also provide 3-D information of hands. There are nevertheless limitations on the dataset size, global viewpoints, hand articulations, and hand shapes in publicly available stereo-based hand pose datasets. To mitigate these limitations and promote further research on hand pose estimation from stereo images, we build a new large-scale binocular hand pose dataset called THU-Bi-Hand, offering a new perspective for fingertip localization. In the THU-Bi-Hand dataset, there are 447k pairs of stereo images of different hand shapes from ten subjects with accurate 3-D location annotations of the wrist and five fingertips. Captured with minimal restriction on the range of hand motion, the dataset covers a large global viewpoint space and hand articulation space. To better present the performance of fingertip localization on THU-Bi-Hand, we propose a novel scheme termed bi-stream pose-guided region ensemble network (Bi-Pose-REN). It extracts more representative feature regions around joints in the feature maps under the guidance of the previously estimated pose. The feature regions are integrated hierarchically according to the topology of hand joints to regress a refined hand pose. Bi-Pose-REN and several existing methods are evaluated on THU-Bi-Hand so that benchmarks are provided for further research. Experimental results show that our Bi-Pose-REN has achieved the best performance on THU-Bi-Hand.

In this article, we establish a novel separating hyperplane classification (SHC) framework to unify three nearest-class-model methods for high-dimensional data: the nearest subspace method (NSM), the nearest convex hull method (NCHM), and the nearest convex cone method (NCCM). Nearest-class-model methods are an important paradigm for the classification of high-dimensional data. We first introduce the three nearest-class-model methods and then conduct dual analysis for theoretically investigating them, to understand deeply their underlying classification mechanisms. A new theorem for the dual analysis of NCCM is proposed in this article by discovering the relationship between a convex cone and its polar cone. We then establish the new SHC framework to unify the nearest-class-model methods based on the theoretical results. One important application of this new SHC framework is to help explain empirical classification results: why one class model has a better performance than others on certain data sets. Finally, we propose a new nearest-class-model method, the soft NCCM, under the novel SHC framework to solve the overlapping class model problem. For illustrative purposes, we empirically demonstrate the significance of our SHC framework and the soft NCCM through two types of typical real-world high-dimensional data: the spectroscopic data and the face image data.

For most of the non-Gaussian statistical models, the data being modeled represent strongly structured properties, such as scalar data with bounded support (e.g., beta distribution), vector data with unit length (e.g., Dirichlet distribution), and vector data with positive elements (e.g., generalized inverted Dirichlet distribution). In practical implementations of non-Gaussian statistical models, it is infeasible to find an analytically tractable solution to estimating the posterior distributions of the parameters. Variational inference (VI) is a widely used framework in Bayesian estimation. Recently, an improved framework, namely, the extended VI (EVI), has been introduced and applied successfully to a number of non-Gaussian statistical models. EVI derives analytically tractable solutions by introducing lower bound approximations to the variational objective function. In this paper, we compare two approximation strategies, namely, the multiple lower bounds (MLBs) approximation and the single lower bound (SLB) approximation, which can be applied to carry out the EVI. For implementation, two different conditions, the weak and the strong conditions, are discussed. Convergence of the EVI depends on the selection of the lower bound, regardless of the choice of weak or strong condition. We also discuss the convergence properties to clarify the differences between MLB and SLB. Extensive comparisons are made based on some EVI-based non-Gaussian statistical models. Theoretical analysis is conducted to demonstrate the differences between the weak and strong conditions. Experimental results based on real data show advantages of the SLB approximation over the MLB approximation.

Imbalanced learning is a challenging task in machine learning, faced by practitioners, and intensively investigated by researchers from a wide range of communities. However, as pointed out in the book titled “Imbalanced Learning: Foundations, Algorithms, and Applications” and collectively authored by experts in the field, many if not most of the approaches to imbalanced learning are heuristic and ad hoc in nature, hence leaving many questions unanswered. To fill this gap, the aim of this Special Issue is to collect recent research works that focus on the theory, methodology, and applications of imbalanced learning. After carefully reviewing a large number of submissions, we selected 15 works to be included in this Special Issue. These works can be roughly categorized into three types: deep-learning-based methods (6), methods based on other machine-learning paradigms (7), and empirical comparative studies (2).

In this paper, we propose novel strategies for neutral vector variable decorrelation. Two fundamental invertible transformations, namely, serial nonlinear transformation and parallel nonlinear transformation, are proposed to carry out the decorrelation. For a neutral vector variable, which is not multivariate-Gaussian distributed, the conventional principal component analysis cannot yield mutually independent scalar variables. With the two proposed transformations, a highly negatively correlated neutral vector can be transformed to a set of mutually independent scalar variables with the same degrees of freedom. We also evaluate the decorrelation performances for the vectors generated from a single Dirichlet distribution and a mixture of Dirichlet distributions. The mutual independence is verified with the distance correlation measurement. The advantages of the proposed decorrelation strategies are intensively studied and demonstrated with synthesized data and practical application evaluations.