Shengchuan Zhang

PAR3D reveals that integrating part-aware representations can dramatically enhance 3D scene understanding, outperforming traditional object-centric models.

Multimodal models stumble badly on low-resource Southeast Asian languages, as revealed by the new SEA-Vision benchmark for document and scene text understanding.

Achieve state-of-the-art zero-shot camouflaged object segmentation by intelligently combining visual features, SAM, and MLLMs to overcome the limitations of relying solely on MLLMs for object discovery.