Reproduce “BlazeVGGT: 13x Faster VGGT at 2,000 Frames via Camera Token without Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shujun Yao, Junhui Lin, Congying Han, Tiande Guo, Ziwen Liu, Bonan Li
BlazeVGGT accelerates long-sequence reconstruction by selecting fewer frames as cross-view evidence while retaining every frame for predictions. It uses camera-token relations within VGGT to select complementary views, then supplies their tokens as keys and values for all-frame queries. On ScanNet-50 with 2,000 views, inference falls from 1,325.1 to 100.1 seconds and Chamfer Distance improves from 0.4184 to 0.3282; camera-pose metrics are not uniformly better.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.