Visual odometry systems face a trade-off: classical sparse methods output poses at high frequency but cannot reconstruct dense geometry, while recent feed-forward models produce dense maps but typically operate on keyframes, yielding sparse pose estimates. The authors of HyVGGT-VO propose a hybrid framework that combines both approaches in a single system.

HyVGGT-VO uses an adaptive frontend that switches between traditional optical flow and the VGGT model depending on conditions, and a hierarchical optimization step that jointly refines the VO poses and the scale of VGGT predictions. The authors say this is the first work to tightly couple a classical VO frontend with VGGT.

Compared with VGGT-SLAM 2.0, the system reduces average absolute trajectory error by about 91% on EuRoC MAV and 20% on TUM RGB-D, and lowers relative translation and rotation errors by 46% and 92% on KITTI. It runs at 16.1 Hz for pose output and 3.2 Hz for dense-mapping updates, using 6.5 GB of memory on an 8 GB laptop.