Search

Human pose estimation is a fundamental and challenging task in the field of computer vision. Hard scenarios, such as occlusion and background confusion, set a great challenge for high-level feature representation because both detailed and multi-scale context must be correctly reasoned. In this paper, we propose a structure-context complementary network (SCC-Net) characterized by the complementarity between a pixel-wise enhanced attention mechanism and atrous convolution-based module. The proposed cross-coordinate attention bottleneck (CCAB) aims to utilize a cross-guide mechanism to promote the robustness of the existing coordinate attention module (CAM) for the background impact. As a complementary module for CCAB, waterfall residual atrous pooling (WRAP) is proposed to refine structure consistency by generating multi-scale features without the feature sparse defect of atrous-based methods. We evaluate our proposed modules and holistic SCC-Net on the COCO and MPII benchmark datasets. Ablation experiments demonstrate that our proposed modules can efficiently boost the performance of body joint detection. Competitive performance is also achieved by our holistic SCC-Net compared to other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Existing methods mainly assign pseudo labels by clustering algorithms to solve the problems in domain adaptive pedestrian re-identification caused by unlabeled target domain data and scene style differences between the source and the target domains. But clustering algorithms with single network may generate incorrect noisy pseudo labels which affects the effectiveness of domain adaptation. To address this problem, the dual-branch teacher-student networks uses two networks with almost the same structures to assign soft pseudo labels. However, the soft pseudo labels assigned are largely the same because of the similar structure of the dual networks, which may lead to the same result with single-network training. Therefore, this paper proposes the discrepant mutual learning fusion network to improve the performance of unsupervised domain adaptive person re-identification by both increasing the difference between dual networks and enhancing their feature expressiveness. Firstly, this paper proposes the discrepant dual-branch network (DDNet) to mine the global and local features of the network by constructing two branches with different depths, and constructs the feature random scaling (FRS) module to further enhance the diversity of the extracted feature. Secondly, the feature fusion (FF) module is built to fuse the discrepant features generated by DDNet, and achieve mutual supervise learning of the fusion classification results, which enhances the ability of network feature expression. Experiments show that the proposed method outperforms most classical domain adaptive methods in recognition accuracy. [ABSTRACT FROM AUTHOR]

Occluded person re-identification (ReID) aims to retrieve images of the same person with occlusion. The key is excavating discriminative features from the nonoccluded regions to achieve better matching. To this end, we propose a novel Differential Attention Siamese Network (DASN) for occluded person ReID. Specifically, the proposed DASN method includes a siamese network module and a Differential Occlusion Weighted Aggregation (DOWA) module. First, a shared two-branch framework with the original nonoccluded image and the corresponding occluded image generated by the random erasing operation as inputs is adopted for shared feature representation. Then, the DOWA module is proposed to obtain the difference features in nonoccluded regions by first subtracting the feature map of the occluded image from the original image to obtain features corresponding to the occluded regions, and subsequently subtracting these features from those corresponding to the original image. During this process, the attention mechanism is integrated to alleviate the exploration inaccuracy of the features of nonoccluded areas. In the training process, multi-dimensional loss functions are integrated as the mentor for more effective learning supervision. Finally, an occluded person ReID model with better feature representation capability is produced. Extensive experiments are conducted on five publicly released databases for ReID. The results demonstrate that the proposed method outperforms the state-of-the-art methods in the occluded person ReID task and another two ReID tasks. [ABSTRACT FROM AUTHOR]

Skeleton-based action recognition has recently attracted widespread attention in the field of computer vision. Previous studies on skeleton-based action recognition are susceptible to interferences from redundant video frames in judging complex actions but ignore the fact that the spatial-temporal features of different actions are extremely different. To solve these problems, we propose a triplet attention multiple spacetime-semantic graph convolutional network for skeleton-based action recognition (AM-GCN), which can not only capture the multiple spacetime-semantic feature from the video images to avoid limited information diversity from single-layer feature representation but can also improve the generalization ability of the network. We also present the triplet attention mechanism to apply an attention mechanism to different key points, key channels, and key frames of the actions, improving the accuracy and interpretability of the judgement of complex actions. In addition, different kinds of spacetime-semantic feature information are combined through the proposed fusion decision for comprehensive prediction in order to improve the robustness of the algorithm. We validate AM-GCN with two standard datasets, NTU-RGBD and Kinetics, and compare it with other mainstream models. The results show that the proposed model achieves tremendous improvement. [ABSTRACT FROM AUTHOR]