Jie Feng

Stop guessing affordances from static scenes: A3R's agentic approach leverages cross-dimensional evidence acquisition to significantly outperform one-shot methods in complex 3D environments.

Generate detailed 3D indoor scenes from short text descriptions with SDesc3D, a framework that leverages multi-view structural priors and regional functionality to overcome the limitations of explicit semantic cues.

Forget expensive video diffusion: Resonance4D unlocks high-fidelity 4D dynamic simulations by cleverly supervising motion in the frequency domain.

By recognizing that CLIP features aren't monolithic, DR-Seg unlocks targeted structural enhancements that dramatically improve open-vocabulary remote sensing segmentation.

VLMs struggle to connect the dots between dynamic drone footage and satellite imagery, highlighting a critical gap in their spatial reasoning abilities.