Fudan's F2G-Pose directly regresses 6D pose from one RGB-D frame, 7x faster

机器之心 · wechat · 2026-11-22

F2G-Pose (NeurIPS 2026, Fudan + Shanghai Innovation Institute) takes a direct-regression route to category-level 6D pose estimation: one forward pass over a segmented RGB-D frame predicts 6D pose, true size, and camera-frame shape completion — no diffusion sampling or candidate filtering. It lifts frozen RADIOv2.5 foundation features onto point clouds via camera intrinsics, encodes local geometry with DGCNN, aggregates with an MoE Transformer (8 experts, top-2), and uses shape completion as geometric supervision. The paper also exposes GenPose++'s selection bottleneck: 79.3% of instances have a <5° rotation candidate, but only 48.3% survive filtering.

Original post →

More from Embodied

Embodied channel →