Quanzhou / Fuzhou, China
I study what video-language models actually need to see — and how to make them pay for only that.
Teaching models to count repetitions with no count labels at all. Real motion speeds up and slows down mid-video, which quietly breaks the fixed-period assumption almost every prior method rests on. I model period adaptively across multiple temporal scales, so the structure of repetition is learned rather than assumed.
Skeleton representations inherit whatever the camera happened to be doing. I use gravity as a stable external reference to decouple viewpoint from motion semantics, and model anomaly as a relational property — because what makes one person's behaviour abnormal is usually the people around them.
Most VideoLLM pipelines sample frames densely or on a fixed budget, with no evidence that the frames are sufficient. I measure minimum sufficient frame sets directly, then route uncertain examples to larger budgets — spending compute where the model is actually unsure, and reporting plainly where the strategy fails to transfer.
A self-supervised framework that adapts to changing action periods instead of assuming a fixed one, combining multi-scale cycle consistency with an action-similarity objective and a multi-expert consensus counting stage. Reaches 0.129 MAE / 0.842 OBO on RepCount and 0.098 MAE / 0.879 OBO on Countix — outperforming prior supervised baselines while using no count annotations.
Resource-aware on-device repetition sensing for mobile edge devices
First author · Sole student author
Semantic-guided representation decoupling for self-supervised repetitive counting
Student co-author · Sole student author
View-robust skeleton action recognition
First author · Sole student author
Multi-person skeleton anomaly detection
First author · Sole student author
Temporal cycle consistency with dynamic thresholding for robust counting
Student co-author
Global multi-geometric feature learning for repetitive counting
Student co-author
Flexible graph-based modeling with adaptive node interaction
Student co-author
Titles are withheld while under review, in line with double-blind policy. Full details available on request or in my CV.
Prior repetitive-counting methods assume a fixed period per video. Real actions accelerate and decelerate mid-sequence, so that assumption fails exactly where counting matters most — and most methods additionally require dense count annotations.
Period-adaptive multi-scale temporal cycle consistency paired with an action-similarity objective, learning repetition structure from unlabeled skeleton sequences, and a multi-expert consensus stage that turns learned structure into a count.
Published at CVPR 2026 Findings and awarded a CVPR Compute Gold Star. I led problem formulation, method design, the full PyTorch implementation, all experiments and ablations, and the manuscript.
VideoLLM evaluation samples frames densely or on a fixed budget. Nobody checks whether those frames are sufficient — so compute is spent uniformly on questions of wildly different difficulty.
Measured minimum sufficient frame sets across 1,336 LongVideoBench validation examples on Qwen2.5-VL and LLaVA-Video, then built visual-token masking, answer-option scoring, and nested confidence routing that escalates only uncertain examples to larger frame budgets.
Shipped as a reproducible package: model adapters, automated tests, CI, sanitized experiment artifacts, and full data/licensing documentation. The negative transfer result is reported as prominently as the positive one — a cross-model boundary is a finding, not a failure.
Skeleton action representations are entangled with the camera coordinate frame, so recognition degrades under viewpoint change. And multi-person anomaly detection typically scores each person in isolation, discarding the social context that defines abnormality.
Gravity-aware view decoupling with adaptive feature modulation to strip camera dependence while preserving motion semantics; and gated, dynamic inter-person dependency modeling so that social context can shape the anomaly score.
The two methodological directions led to two AAAI 2027 submissions, both as first author.
Before research, I shipped. Nearly two years as a product engineer on enterprise server platforms, then a product-management stint on a bank financial-guarantee system, and a student software company I ran through my undergraduate degree. That background is why my research ships with tests, CI, and a cost model rather than a folder of scripts.
He independently identified the bottleneck in variable-speed action sequences, conceived the mitigation strategy, and drove the implementation to completion — demonstrating the intellectual autonomy essential for doctoral research.
Among the master’s researchers I have worked with in recent years, he stands out for his combination of initiative, follow-through, and consideration for colleagues.
Across these projects, he has demonstrated the full cycle of problem definition, implementation, deployment, evaluation, and revision.
He is willing to resolve infrastructure issues that affect the wider group, even when they are not part of his own immediate experiments.
M.Eng. candidate in Artificial Intelligence, jointly trained at the Quanzhou Institute of Equipment Manufacturing, Chinese Academy of Sciences. My research asks how far video understanding can go without labels and without waste: self-supervised temporal modeling that learns repetition structure on its own, and inference strategies that spend compute only where a model actually needs it. Right now I'm most interested in adaptive visual-token allocation, efficient multimodal reasoning, and whether efficiency gains actually transfer across video-language architectures under realistic compute budgets. Before returning to research I spent nearly two years as a product engineer shipping server platforms at enterprise scale — that engineering instinct still shapes how I build: reproducible pipelines, measured cost, honest negative results.
M.Eng. in Artificial Intelligence
College of Mechanical and Electrical Engineering · Fuzhou / Quanzhou, China
Jointly trained at the Quanzhou Institute of Equipment Manufacturing, Haixi Institute, Chinese Academy of Sciences
Thesis: Research on Weakly Supervised Repetitive Action Counting Methods Based on Action Similarity. Advisors: Prof. Jun Li and Assoc. Prof. Qiming Li. Ranked top 10% in comprehensive assessment.
B.Eng. in Software Engineering
School of Computer Science · Wuhan, China
Ranked top 10% in comprehensive assessment. Multiple academic scholarships; provincial awards in the "Internet+" and "Challenge Cup" innovation competitions. Led an on-campus software venture throughout the degree.
Reviewer for Image and Vision Computing, 2026.
Mentored capstone thesis students through problem scoping, method selection, experiment planning, and technical writing.
Supervised interns on research software development and conference-paper preparation — task decomposition, coding milestones, experimental reporting, manuscript feedback.
Maintain shared GPU servers and coordinate compute allocation across the group, keeping team workflows reproducible.
I'm not setting out to change the world. I would like to move one field — a little, and in the right direction — and leave behind work that other people can pick up and build on. That is the whole reason I want to do a Ph.D.
I read every email. If you are a prospective advisor, a collaborator, or simply curious about the work, write to me directly.
Applying for Ph.D. positions starting 2027 · Open to research collaborations