Han Yi

SVI-Bench is a unified benchmark for evaluating the full cognitive stack of video intelligence — perception, reasoning, simulation, and agency. Built on 35,000 hours of broadcast basketball, soccer, and hockey footage, it shows that while models handle perception well, reasoning, simulation, and agency are where current systems break down.

We introduce ExAct, a video-language benchmark for expert-level analysis of skilled human actions. It contains over 3,500 expert-curated video QA pairs across domains like sports, cooking, and music. Our benchmark reveals a significant performance gap between state-of-the-art VLMs and human experts, highlighting the need for models with a more nuanced understanding of complex human skills.

A progressive strategy that learns text-to-3D in a coarse-to-fine manner.

Proposed a novel motion deblurring framework using Blur Space Disentangled Network (BSDNet) and Hierarchical Scale-recurrent Deblurring Network (HSDNet) to effectively address real-world blur, achieving state-of-the-art results.

Developed a recurrent single-image deraindrop approach utilizing luminance priors and contextual feature aggregation, achieving superior performance in restoring color and texture consistency.