Existing approaches to multi-person pose tracking often suffer from low-confidence detections due to inter-instance and intra-instance occlusions, as well as non-canonical poses. In this work, we propose a novel solution by addressing two critical aspects: incomplete joint temporal dependencies and spatio-temporal voxelization. First, we introduce a method for extracting hierarchical relationships between joints based on human dynamics, enabling the model to reason about occlusions within the spatial topology of the human body. This hierarchical approach tackles incomplete joint visibility by leveraging the interdependencies between joints in both space and time. Second, we present a spatio-temporal occupancy network for multi-person pose tracking. By stacking 2D pose data over time to create a spatio-temporal voxel grid, the model captures temporal relationships between instances and joints, enhancing spatio-temporal correlations and learning keypoint distributions under occlusions or non-canonical poses. Extensive experiments on the PoseTrack2017, PoseTrack2018, and PoseTrack21 dataset demonstrate that our method improves multi-person pose tracking performance, achieving state-of-the-art mAP.

Robust multi-object tracking (MOT) is hindered by the instability of appearance features under visual corruptions such as occlusion and motion blur. These perturbations introduce high-variance noise into feature trajectories, weakening temporal representations and causing identity switches. We address this challenge by learning more stable appearance representations under feature corruption. To this end, we propose the Causal Interaction Module (CIM), a causal architecture that follows a filter then reconstruct design for online tracking. A temporal filtering stage summarizes the historical feature trajectory into a stable anchor, and a contextual enhancement stage uses that anchor to refine frame-level features before association. Integrated into standard trackers, CIM improves association robustness while preserving the host tracking formulation. Experiments on multiple MOT benchmarks and corruption stress tests show consistent gains, especially on association-related metrics.

In monocular 3D human pose estimation, target motions are generally stable and continuous, which indicates that joint velocity can provide valuable information for better estimation. Therefore, it is critical to learn the joint motion trajectory and spatio-temporal information from velocity. Previous works have shown that Transformers are effective in capturing the relationship between tokens. However, in practice, only 2D position is available and 3D velocity has not been explicitly used as a model input. To address this challenge, we propose TMT (Two-step Mixed-Training strategy), a transformer-based approach that effectively incorporates 3D velocity into the input vector during training, allowing for better learning of relevant features in the shallow layers. Extensive experiments demonstrate that TMT significantly improves the performance of state-of-the-art models, such as MixSTE, MHFormer, and PoseFomer, on two datasets: Human3.6M and MPI-INF-3DHP. TMT outperforms the state-of-the-art approach by up to 13.8% on the Human3.6M dataset.

It is intuitive that NLP tasks for logographic languages like Chinese should benefit from the use of the glyph information in those languages. However, due to the lack of rich pictographic evidence in glyphs and the weak generalization ability of standard computer vision models on character data, an effective way to utilize the glyph information remains to be found. In this paper, we address this gap by presenting Glyce, the glyph-vectors for Chinese character representations. We make three major innovations: (1) We use historical Chinese scripts (e.g., bronzeware script, seal script, traditional Chinese, etc) to enrich the pictographic evidence in characters; (2) We design CNN structures (called tianzege-CNN) tailored to Chinese character image processing; and (3) We use image-classification as an auxiliary task in a multi-task learning setup to increase the model's ability to generalize. We show that glyph-based models are able to consistently outperform word/char ID-based models in a wide range of Chinese NLP tasks. When combing with BERT, we are able to set new state-of-the-art results for a variety of Chinese NLP tasks, including language modeling, tagging (NER, CWS, POS), sentence pair classification (BQ, LCQMC, XNLI, NLPCC-DBQA), single sentence classification tasks (ChnSentiCorp, the Fudan corpus, iFeng), dependency parsing, and semantic role labeling. For example, the proposed model achieves an F1 score of 81.6 on the OntoNotes dataset of NER, +1.5 over BERT; it achieves an almost perfect accuracy of 99.8\% on the the Fudan corpus for text classification.

Most of the Neural Machine Translation (NMT) models are based on the sequence-tosequence (Seq2Seq) model with an encoderdecoder framework equipped with the attention mechanism.However, the conventional attention mechanism treats the decoding at each time step equally with the same matrix, which is problematic since the softness of the attention for different types of words (e.g.content words and function words) should differ.Therefore, we propose a new model with a mechanism called Self-Adaptive Control of Temperature (SACT) to control the softness of attention by means of an attention temperature.Experimental results on the Chinese-English translation and English-Vietnamese translation demonstrate that our model outperforms the baseline models, and the analysis and the case study show that our model can attend to the most relevant elements in the source-side contexts and generate the translation of high quality.