IT之家 · 智能时代T2·78 pts
- The dual-loop Transformer decouples cross-frame feature aggregation from iterative camera-state refinement, feeding pose and intrinsics back into the next loop
- The training platform aggregates about 2.97 million sequences and 962 million frames spanning simulated and real, dynamic and static scenes
- It reports state-of-the-art results on SpatialBench, ClassicBench and VRBench