Selected research projects
We propose a training-free diffusion guidance method that decomposes noise in a stage-aware manner to improve generation quality
We propose a grounding decoder that leverages expression and geometric cues to localize objects described in natural language within 3D scenes
We propose a dual-path temporal decoder that jointly models detection and association for end-to-end multi-object tracking
We propose an event-guided framework that transfers deformable features and refines dual memories to segment objects in low-light video
We propose a geometry relation-aware encoder that models spatial relations between objects for online 3D multi-object tracking
We propose a vision-centric scene completion model with a scene-adaptive decoder and view projection that accounts for occluded regions
We propose a luminance-aware color transform that corrects images captured under multiple exposure settings
We propose a density estimation method based on local connectivity for face clustering
Detailed project pages for each topic will be added soon.