We introduce LoG‑VGGT, a memory‑efficient framework for long‑sequence 3D reconstruction that balances local temporal modeling with global camera consistency. Traditional approaches rely on full‑sequence attention, causing memory usage to grow linearly with sequence length. LoG‑VGGT inserts cross‑window attention into a small subset of Transformer blocks, allowing information to flow between adjacent temporal windows while keeping memory bounded.\
To mitigate long‑term pose drift, we add a global camera consistency refinement module. Camera tokens interact with compact register tokens via cross‑attention, enforcing scene‑level constraints across the entire sequence. This joint optimization of camera representations markedly improves pose stability without the high cost of sequence‑wide attention.\
Extensive experiments on several long‑sequence benchmarks show that LoG‑VGGT achieves higher depth accuracy and more robust camera pose estimation, while delivering competitive streaming reconstruction performance.\
Review