F2G-Pose estimates category-level RGB-D object pose in one pass
Original title: F2G-Pose: Geometry-Aware Foundation Feature Lifting for Direct RGB-D Category-Level Object Pose Estimation
It reports state-of-the-art accuracy on SOPE and ROPE at 28 FPS for pose-only inference, with code and weights released.
F2G-Pose predicts category-level pose, metric size, and a camera-frame completed shape from one RGB-D view without CAD models or reference views. It lifts frozen RADIO features onto partial point clouds, tokenizes geometry with DGCNN, and uses a Mixture-of-Experts Transformer. The pose network was trained only on synthetic SOPE annotations.
- Pose-only inference runs at 28 FPS.
- It also reports results on HouseCat6D and real-object transfer on HANDAL.
- The RADIO backbone stays frozen and was pretrained externally.
- Code and pretrained models are available on the project page.
From the original article
1 min read at huggingface.co
This paper studies category-level RGB-D object pose estimation, which recovers an object's 3D rotation and translation without instance-specific CAD models or reference views. This task is highly ambiguous due to intra-class variations, symmetries, and partial observations.
By Jinyu Zhang, Haitao Lin, Jiashu Hou · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.