Disentangled Gaussian Transformers for Animatable Humans
Soham Dasgupta
Reconstructing animatable 3D human avatars from casual inputs — a single image or monocular video — remains a difficult problem. Recent feed-forward models such as LHM regress Gaussian avatars directly from images but suffer from spatial misalignment with the input and over-smoothed textures, both stemming from a reliance on linear blend skinning (LBS) to explain inherently non-rigid motion. We propose Disentangled Gaussian Transformers (DGT), a transformer-based framework that produces animatable 3D Gaussian avatars from variable inputs while separating canonical geometry from pose-dependent deformation. DGT uses a multi-stream point transformer: a canonical T-stream refines a shared T-pose avatar across all inputs, and per-image P-streams predict residual deformations in the posed space. A Foreground Point Projection module injects localized 2D evidence into the 3D Gaussians, correcting pixel-level misalignment without compromising geometric consistency.
Cooperative Foundation Models For Object Detection Active Learning
Zizheng Zhou
Object detection requires costly bounding-box annotations. Active learning reduces the number of images to label, but selected samples still need manual box drawing, which remains slow and expensive. We propose CoMODAL, an active learning framework that uses multiple foundation models cooperatively to reduce annotation effort. A vision-language model committee first mines reliable negative images without human labeling. Then, a dual-objective query strategy selects diverse high-uncertainty samples and likely high-confidence false positives using DINOv3 patch embeddings. Finally, foundation models propose candidate boxes, so the human oracle only needs to approve or reject them instead of drawing boxes from scratch. This turns object detection active learning into a more efficient, lower-cost annotation pipeline.