Paper: Efficiently Reconstructing Dynamic Scenes One D4RT at a Time Authors: Chuhan Zhang, Guillaume Le Moing, et al. arXiv: 2512.08924 DOI: 10.48550/arXiv.2512.08924 Year: 2025
D4RT introduces a single feedforward transformer that jointly estimates dense depth, 3D point tracking, and camera pose from monocular video. The architecture decouples a global ViT encoder from a lightweight cross-attention decoder queried by (u, v, t_src, t_tgt, t_cam) tuples, enabling arbitrary spatiotemporal probing at 200+ FPS.