From July to September 2026 I was a research intern on large language models for electronic design automation at the Eastern Institute of Technology, Ningbo, and an NLP research intern with the Minnesota NLP Group, where I built tools that record and compare how people and computer-use agents work through the same browser tasks. Before that I was a research assistant at The Chinese University of Hong Kong (uncertainty-aware LLMs for recommendation) and at Zhangjiang Laboratory (the IC design flow), and a teaching assistant for calculus and logic at Eindhoven University of Technology.
I also like fixing problems in the open-source tools I use; 20 of my pull requests have been merged into projects such as OpenRAM, CharLib, Power Grid Model, PISA, and LensKit. This fall I submitted TempoRAC to ICASSP 2027 as first author (equal contribution), and two co-authored manuscripts are under review at IEEE Transactions on Geoscience and Remote Sensing and at Signal, Image and Video Processing. My first paper, written in high school, analyzed PM2.5 exposure and respiratory disease on synthetic data and appeared in Theoretical and Natural Science; it was an independent project supervised remotely by Prof. Soummya Kar (Carnegie Mellon University) and Assoc. Prof. Qiang Gao (Southeast University).
Three new manuscripts are under review. I submitted TempoRAC: Identity-Indexed Local Tempo Routing for Multi-Person Repetitive Action Counting to ICASSP 2027 with Jiayu Zhu and Xu Yan (equal contribution). I am also a co-author of VCH-SegFormer for Street-View Vegetation Mapping and Assessing the Associations Between Urban Vegetation and PM2.5, submitted to IEEE Transactions on Geoscience and Remote Sensing, and RACER-T: Temporal Density Modeling for Pose-Based Repetitive Action Counting, submitted to Signal, Image and Video Processing.
Six of my pull requests were merged upstream between August 4 and September 8, 2026: Power Grid Model #1516, #1518, and #1540; LensKit #1209; M* #235; and PISA #641. The accepted changes cover update validation and sensor semantics, performance documentation, recommender warnings, media decoding and error handling, and retrieval type safety.
Multi-person repetitive action counting (MRAC) is crucial for human-centric applications such as fitness analysis and sports monitoring. However, existing methods suffer from asynchronous motions, varying tempos, and unreliable counting under occlusions. To address these challenges, we propose TempoRAC, a pose-driven framework for adaptive person-wise repetition counting. TempoRAC introduces an identity-conditioned pose encoder to capture individual motion patterns, and a tempo-aware routing module that dynamically selects fast, medium, and slow experts for modeling non-stationary repetitions. Furthermore, a normalized overlap-add fusion strategy integrates window responses and enables consistent identity-wise decoding without redundant counting. Extensive experiments demonstrate that TempoRAC achieves robust and accurate person-wise repetition counting under diverse motion conditions.
@unpublished{yan2026temporac,title={{TempoRAC}: Identity-Indexed Local Tempo Routing for Multi-Person Repetitive Action Counting},author={Yan, Weiping and Zhu, Jiayu and Yan, Xu},year={2026},month=sep,note={Submitted to {ICASSP} 2027; under review. Equal contribution by all three authors.},}
Urban vegetation can influence fine particulate matter (PM2.5) concentrations through particle deposition and changes in atmospheric dispersion, while observed associations also reflect land-use patterns. However, assessments of urban vegetation–PM2.5 associations rely predominantly on satellite-derived spectral indices, while street-level greenness remains comparatively underexamined. We developed VCH-SegFormer, a semantic segmentation model that extends SegFormer with a Vertical Context Gate and a Hierarchical Vegetation Head to distinguish trees from low vegetation. With a MiT-B5 encoder, the model achieved a mean intersection over union (mIoU) of 79.3% and a vegetation-fraction mean absolute error of 0.032 on ADE20K-6, and an mIoU of 86.3% on Cityscapes. The trained models were applied to 430,676 street-view images in Guangzhou, with the resulting predictions aggregated into one-kilometer cells to compare street-view green view index (GVI) and satellite normalized difference vegetation index (NDVI) in relation to modeled PM2.5. Vegetation GVI and NDVI showed Pearson correlations of −0.242 and −0.419, respectively, with modeled PM2.5. In a complementary U.S. analysis of 430 monitors, a one-standard-deviation increase in NDVI was associated with a 0.29 μg/m³ lower measured PM2.5 concentration. In a separate sample of 55 monitors, a one-standard-deviation increase in street-view GVI was associated with a 0.70 μg/m³ lower measured PM2.5 concentration after adjustment for NDVI and other covariates. The approach combines street-view vegetation segmentation with satellite observations to examine how eye-level and overhead greenness relate to PM2.5.
@unpublished{fei2026vchsegformer,title={{VCH-SegFormer} for Street-View Vegetation Mapping and Assessing the Associations Between Urban Vegetation and {PM2.5}},author={Fei, Xiaoyan and Yan, Weiping and Zhu, Qi and Yan, Xu and Wu, Zherong},year={2026},month=oct,note={Submitted to {IEEE} Transactions on Geoscience and Remote Sensing; under review.},}
@unpublished{wang2026racert,title={{RACER-T}: Temporal Density Modeling for Pose-Based Repetitive Action Counting},author={Wang, Shuoran and Yan, Weiping and Wang, Xiang},year={2026},month=oct,note={Submitted to Signal, Image and Video Processing (Springer Nature); under review. Opted in to the In Review preprint service on Research Square.},}
Reconstructing sharp and animatable 3D human avatars from motion-blurred videos remains challenging under partial occlusion, where unreliable SMPL estimates can propagate pose errors into Gaussian deformation. To address this problem, we propose OAM-GA, an occlusion-robust framework for motion-aware Gaussian avatar reconstruction. OAM-GA estimates joint-wise reliability and restores missing motion using visible temporal context. Specifically, a Bidirectional Wavelet Masked Motion Completion module decomposes motion features into low- and high-frequency components and reconstructs contiguous occluded intervals from observations before and after the masked region. Clean–occluded contrastive consistency learning further encourages occlusion-invariant motion representations, while reliability-guided residual refinement selectively corrects uncertain joints and preserves reliable ones. Experiments on BlurZJU and BS-Human show that OAM-GA improves overall reconstruction robustness and perceptual quality under diverse occlusion patterns while maintaining the rendering fidelity of the original baseline.
@unpublished{wang2026oamga,title={{OAM-GA}: Reliability-Guided Motion Completion for Occlusion-RobustGaussian Avatars},author={Wang, Xiang and Yan, Xu and Pei, Letian and Yan, Weiping},year={2026},note={Submitted to {DAI} 2026; under review, decision pending, and not yet accepted or published.},url={https://openreview.net/forum?id=UV2UJ4VHf7},}
@article{yan2023pm25,title={Effect of {PM2.5} air pollution on the incidence of respiratory diseases: A {Python}-based data analysis},author={Yan, Weiping},journal={Theoretical and Natural Science},volume={8},number={1},pages={70--75},year={2023},month=nov,publisher={EWA Publishing},doi={10.54254/2753-8818/8/20240361},url={https://tns.ewapub.com/article/view/6279},}