Event

Proposes Hybrid Dense Visual Odometry Framework HyVGGT-VO

1 reports · 1 sources

Junxiang Pang and other researchers propose the HyVGGT-VO framework, which tightly couples a traditional sparse VO frontend with the feedforward model VGGT, and is claimed to be the first work to tightly couple a traditional VO frontend with VGGT. Its adaptive hybrid tracking frontend dynamically switches between traditional optical flow and the VGGT model, and a hierarchical optimization framework jointly refines VO poses and the scale predicted by VGGT to ensure global scale consistency. Under the odometry setting, the authors claim that compared with VGGT-SLAM 2.0, HyVGGT-VO reduces average ATE by about 91% on EuRoC MAV and by 20% on TUM RGB-D, and reduces average relative translation and rotation errors on KITTI by 46% and 92%, respectively; on an 8 GB laptop, it achieves 16.1 Hz pose output and 3.2 Hz dense mapping updates using 6.5 GB of memory.