The authors first analyzed the evolution of camera tokens in VGGT across attention layers, revealing a topology-first, geometry-later progression: shallow tokens preserve local view neighborhoods, deeper tokens encode global inter-view geometry.
BlazeVGGT speeds up VGGT 13x by selecting representative views without retraining
The method uses camera-token relations in VGGT to identify a compact set of mutually complementary frames for attention, improving accuracy and efficiency at long sequences.
Top University
Shujun Yao · Junhui Lin · Congying Han · Tiande Guo · Ziwen Liu · Bonan Li
University of Chinese Academy of Sciences · University of Maryland · Beijing University of Posts and Telecommunications
Research Digest··3 min read
The authors introduce BlazeVGGT, a training-free plug-in for VGGT that sparsifies cross-view attention at the frame level.
§