SQA3D

Researchers introduced SpaR3D-MoE, a framework that enhances multimodal large language models with adaptive 3D spatial reasoning from sparse RGB inputs. It uses a spatiotemporal manifold sampling mech…

Researchers have developed UniMVU, a unified multimodal video understanding framework that uses instruction-aware gating to dynamically balance the importance of different input modalities like video,…