Search

The long-tailed recognition (LTR) is the task of learning high-performance classifiers given extremely imbalanced training samples between categories. Most of the existing works address the problem by either enhancing the features of tail classes or re-balancing the classifiers to reduce the inductive bias. In this paper, we try to look into the root cause of the LTR task, i.e., training samples for each class are greatly imbalanced, and propose a straightforward solution. We split the categories into three groups, i.e., many, medium and few, according to the number of training images. The three groups of categories are separately predicted to reduce the difficulty for classification. This idea naturally arises a new problem of how to assign a given sample to the right class groups? We introduce a mutual exclusive modulator which can estimate the probability of an image belonging to each group. Particularly, the modulator consists of a light-weight module and learned with a mutual exclusive objective. Hence, the output probabilities of the modulator encode the data volume clues of the training dataset. They are further utilized as prior information to guide the prediction of the classifier. We conduct extensive experiments on multiple datasets, e.g., ImageNet-LT, Place-LT and iNaturalist 2018 to evaluate the proposed approach. Our method achieves competitive performance compared to the state-of-the-art benchmarks.

Deblurring can not only provide visually more pleasant pictures and make photography more convenient, but also can improve the performance of objection detection as well as tracking. However, removing dynamic scene blur from images is a non-trivial task as it is difficult to model the non-uniform blur mathematically. Several methods first use single or multiple images to estimate optical flow (which is treated as an approximation of blur kernels) and then adopt non-blind deblurring algorithms to reconstruct the sharp images. However, these methods cannot be trained in an end-to-end manner and are usually computationally expensive. In this paper, we explore optical flow to remove dynamic scene blur by using the multi-scale spatially variant recurrent neural network (RNN). We utilize FlowNets to estimate optical flow from two consecutive images in different scales. The estimated optical flow provides the RNN weights in different scales so that the weights can better help RNNs to remove blur in the feature spaces. Finally, we develop a convolutional neural network (CNN) to restore the sharp images from the deblurred features. Both quantitative and qualitative evaluations on the benchmark datasets demonstrate that the proposed method performs favorably against state-of-the-art algorithms in terms of accuracy, speed, and model size., Comment: accepted by tcsvt

The recently proposed MaskFormer gives a refreshed perspective on the task of semantic segmentation: it shifts from the popular pixel-level classification paradigm to a mask-level classification method. In essence, it generates paired probabilities and masks corresponding to category segments and combines them during inference for the segmentation maps. In our study, we find that per-mask classification decoder on top of a single-scale feature is not effective enough to extract reliable probability or mask. To mine for rich semantic information across the feature pyramid, we propose a transformer-based Pyramid Fusion Transformer (PFT) for per-mask approach semantic segmentation with multi-scale features. The proposed transformer decoder performs cross-attention between the learnable queries and each spatial feature from the feature pyramid in parallel and uses cross-scale inter-query attention to exchange complimentary information. We achieve competitive performance on three widely used semantic segmentation datasets. In particular, on ADE20K validation set, our result with Swin-B backbone surpasses that of MaskFormer's with a much larger Swin-L backbone in both single-scale and multi-scale inference, achieving 54.1 mIoU and 55.7 mIoU respectively. Using a Swin-L backbone, we achieve single-scale 56.1 mIoU and multi-scale 57.4 mIoU, obtaining state-of-the-art performance on the dataset. Extensive experiments on three widely used semantic segmentation datasets verify the effectiveness of our proposed method.

3D human shape and pose estimation is the essential task for human motion analysis, which is widely used in many 3D applications. However, existing methods cannot simultaneously capture the relations at multiple levels, including spatial-temporal level and human joint level. Therefore they fail to make accurate predictions in some hard scenarios when there is cluttered background, occlusion, or extreme pose. To this end, we propose Multi-level Attention Encoder-Decoder Network (MAED), including a Spatial-Temporal Encoder (STE) and a Kinematic Topology Decoder (KTD) to model multi-level attentions in a unified framework. STE consists of a series of cascaded blocks based on Multi-Head Self-Attention, and each block uses two parallel branches to learn spatial and temporal attention respectively. Meanwhile, KTD aims at modeling the joint level attention. It regards pose estimation as a top-down hierarchical process similar to SMPL kinematic tree. With the training set of 3DPW, MAED outperforms previous state-of-the-art methods by 6.2, 7.2, and 2.4 mm of PA-MPJPE on the three widely used benchmarks 3DPW, MPI-INF-3DHP, and Human3.6M respectively. Our code is available at https://github.com/ziniuwan/maed.

Previous methods for skeleton-based gesture recognition mostly arrange the skeleton sequence into a pseudo picture or spatial-temporal graph and apply deep Convolutional Neural Network (CNN) or Graph Convolutional Network (GCN) for feature extraction. Although achieving superior results, these methods have inherent limitations in dynamically capturing local features of interactive hand parts, and the computing efficiency still remains a serious issue. In this work, the self-attention mechanism is introduced to alleviate this problem. Considering the hierarchical structure of hand joints, we propose an efficient hierarchical self-attention network (HAN) for skeleton-based gesture recognition, which is based on pure self-attention without any CNN, RNN or GCN operators. Specifically, the joint self-attention module is used to capture spatial features of fingers, the finger self-attention module is designed to aggregate features of the whole hand. In terms of temporal features, the temporal self-attention module is utilized to capture the temporal dynamics of the fingers and the entire hand. Finally, these features are fused by the fusion self-attention module for gesture classification. Experiments show that our method achieves competitive results on three gesture recognition datasets with much lower computational complexity., Comment: Under peer review for TCSVT

Sparsity in Deep Neural Networks (DNNs) has been widely studied to compress and accelerate the models on resource-constrained environments. It can be generally categorized into unstructured fine-grained sparsity that zeroes out multiple individual weights distributed across the neural network, and structured coarse-grained sparsity which prunes blocks of sub-networks of a neural network. Fine-grained sparsity can achieve a high compression ratio but is not hardware friendly and hence receives limited speed gains. On the other hand, coarse-grained sparsity cannot concurrently achieve both apparent acceleration on modern GPUs and decent performance. In this paper, we are the first to study training from scratch an N:M fine-grained structured sparse network, which can maintain the advantages of both unstructured fine-grained sparsity and structured coarse-grained sparsity simultaneously on specifically designed GPUs. Specifically, a 2:4 sparse network could achieve 2x speed-up without performance drop on Nvidia A100 GPUs. Furthermore, we propose a novel and effective ingredient, sparse-refined straight-through estimator (SR-STE), to alleviate the negative influence of the approximated gradients computed by vanilla STE during optimization. We also define a metric, Sparse Architecture Divergence (SAD), to measure the sparse network's topology change during the training process. Finally, We justify SR-STE's advantages with SAD and demonstrate the effectiveness of SR-STE by performing comprehensive experiments on various tasks. Source codes and models are available at https://github.com/NM-sparsity/NM-sparsity., Comment: ICLR2021

Both high-level and high-resolution feature representations are of great importance in various visual understanding tasks. To acquire high-resolution feature maps with high-level semantic information, one common strategy is to adopt dilated convolutions in the backbone networks to extract high-resolution feature maps, such as the dilatedFCN-based methods for semantic segmentation. However, due to many convolution operations are conducted on the high-resolution feature maps, such methods have large computational complexity and memory consumption. In this paper, we propose one novel holistically-guided decoder which is introduced to obtain the high-resolution semantic-rich feature maps via the multi-scale features from the encoder. The decoding is achieved via novel holistic codeword generation and codeword assembly operations, which take advantages of both the high-level and low-level features from the encoder features. With the proposed holistically-guided decoder, we implement the EfficientFCN architecture for semantic segmentation and HGD-FPN for object detection and instance segmentation. The EfficientFCN achieves comparable or even better performance than state-of-the-art methods with only 1/3 of their computational costs for semantic segmentation on PASCAL Context, PASCAL VOC, ADE20K datasets. Meanwhile, the proposed HGD-FPN achieves $>2\%$ higher mean Average Precision (mAP) when integrated into several object detection frameworks with ResNet-50 encoding backbones., Comment: arXiv admin note: substantial text overlap with arXiv:2008.10487

Both performance and efficiency are important to semantic segmentation. State-of-the-art semantic segmentation algorithms are mostly based on dilated Fully Convolutional Networks (dilatedFCN), which adopt dilated convolutions in the backbone networks to extract high-resolution feature maps for achieving high-performance segmentation performance. However, due to many convolution operations are conducted on the high-resolution feature maps, such dilatedFCN-based methods result in large computational complexity and memory consumption. To balance the performance and efficiency, there also exist encoder-decoder structures that gradually recover the spatial information by combining multi-level feature maps from the encoder. However, the performances of existing encoder-decoder methods are far from comparable with the dilatedFCN-based methods. In this paper, we propose the EfficientFCN, whose backbone is a common ImageNet pre-trained network without any dilated convolution. A holistically-guided decoder is introduced to obtain the high-resolution semantic-rich feature maps via the multi-scale features from the encoder. The decoding task is converted to novel codebook generation and codeword assembly task, which takes advantages of the high-level and low-level features from the encoder. Such a framework achieves comparable or even better performance than state-of-the-art methods with only 1/3 of the computational cost. Extensive experiments on PASCAL Context, PASCAL VOC, ADE20K validate the effectiveness of the proposed EfficientFCN., Comment: Accepted in ECCV 2020

Long-range contextual information is essential for achieving high-performance semantic segmentation. Previous feature re-weighting methods demonstrate that using global context for re-weighting feature channels can effectively improve the accuracy of semantic segmentation. However, the globally-sharing feature re-weighting vector might not be optimal for regions of different classes in the input image. In this paper, we propose a Context-adaptive Convolution Network (CaC-Net) to predict a spatially-varying feature weighting vector for each spatial location of the semantic feature maps. In CaC-Net, a set of context-adaptive convolution kernels are predicted from the global contextual information in a parameter-efficient manner. When used for convolution with the semantic feature maps, the predicted convolutional kernels can generate the spatially-varying feature weighting factors capturing both global and local contextual information. Comprehensive experimental results show that our CaC-Net achieves superior segmentation performance on three public datasets, PASCAL Context, PASCAL VOC 2012 and ADE20K., Comment: Accepted in ECCV 2020

Artistic style transfer can be thought as a process to generate different versions of abstraction of the original image. However, most of the artistic style transfer operators are not optimized for human faces thus mainly suffers from two undesirable features when applying them to selfies. First, the edges of human faces may unpleasantly deviate from the ones in the original image. Second, the skin color is far from faithful to the original one which is usually problematic in producing quality selfies. In this paper, we take a different approach and formulate this abstraction process as a gradient domain learning problem. We aim to learn a type of abstraction which not only achieves the specified artistic style but also circumvents the two aforementioned drawbacks thus highly applicable to selfie photography. We also show that our method can be directly generalized to videos with high inter-frame consistency. Our method is also robust to non-selfie images, and the generalization to various kinds of real-life scenes is discussed. We will make our code publicly available.

We observed that recent state-of-the-art results on single image human pose estimation were achieved by multi-stage Convolution Neural Networks (CNN). Notwithstanding the superior performance on static images, the application of these models on videos is not only computationally intensive, it also suffers from performance degeneration and flicking. Such suboptimal results are mainly attributed to the inability of imposing sequential geometric consistency, handling severe image quality degradation (e.g. motion blur and occlusion) as well as the inability of capturing the temporal correlation among video frames. In this paper, we proposed a novel recurrent network to tackle these problems. We showed that if we were to impose the weight sharing scheme to the multi-stage CNN, it could be re-written as a Recurrent Neural Network (RNN). This property decouples the relationship among multiple network stages and results in significantly faster speed in invoking the network for videos. It also enables the adoption of Long Short-Term Memory (LSTM) units between video frames. We found such memory augmented RNN is very effective in imposing geometric consistency among frames. It also well handles input quality degradation in videos while successfully stabilizes the sequential outputs. The experiments showed that our approach significantly outperformed current state-of-the-art methods on two large-scale video pose estimation benchmarks. We also explored the memory cells inside the LSTM and provided insights on why such mechanism would benefit the prediction for video-based pose estimations., Comment: Poster in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

Recent advances in visual tracking showed that deep Convolutional Neural Networks (CNN) trained for image classification can be strong feature extractors for discriminative trackers. However, due to the drastic difference between image classification and tracking, extra treatments such as model ensemble and feature engineering must be carried out to bridge the two domains. Such procedures are either time consuming or hard to generalize well across datasets. In this paper we discovered that the internal structure of Region Proposal Network (RPN)'s top layer feature can be utilized for robust visual tracking. We showed that such property has to be unleashed by a novel loss function which simultaneously considers classification accuracy and bounding box quality. Without ensemble and any extra treatment on feature maps, our proposed method achieved state-of-the-art results on several large scale benchmarks including OTB50, OTB100 and VOT2016. We will make our code publicly available.

Most of the recent successful methods in accurate object detection and localization used some variants of R-CNN style two stage Convolutional Neural Networks (CNN) where plausible regions were proposed in the first stage then followed by a second stage for decision refinement. Despite the simplicity of training and the efficiency in deployment, the single stage detection methods have not been as competitive when evaluated in benchmarks consider mAP for high IoU thresholds. In this paper, we proposed a novel single stage end-to-end trainable object detection network to overcome this limitation. We achieved this by introducing Recurrent Rolling Convolution (RRC) architecture over multi-scale feature maps to construct object classifiers and bounding box regressors which are "deep in context". We evaluated our method in the challenging KITTI dataset which measures methods under IoU threshold of 0.7. We showed that with RRC, a single reduced VGG-16 based model already significantly outperformed all the previously published results. At the time this paper was written our models ranked the first in KITTI car detection (the hard level), the first in cyclist detection and the second in pedestrian detection. These results were not reached by the previous single stage methods. The code is publicly available., Comment: CVPR 2017