F2G-Pose

F2G-Pose predicts an object's 6D pose, metric size, and completed shape from aligned RGB, depth, camera intrinsics, and a binary target mask. The model does not require an object-specific CAD model or class label at inference.

Files and setup

• f2g-pose.pt: selected inference checkpoint with its model configuration. SHA-256: bff06bd58a42fd805d8232f9ab7b4612b8eb833edf6b6cccd23e5040a8c9ca18.
• evaluation_results.json: complete SOPE/ROPE test metrics for the selected tensor-identical checkpoint.
• NVIDIA RADIO v2.5-L source and checkpoint are separate downloads, required at runtime.

Follow the installation guide and usage guide. Save f2g-pose.pt in the source checkout's weights/ directory. The bundled ROPE example includes RGB, metric depth, mask, and intrinsics.

Results

Dataset	AUC25	AUC50	AUC75	Mean rotation	Mean translation
SOPE	60.24	44.73	17.69	14.73°	0.824 cm
ROPE	46.75	28.19	6.17	26.99°	1.181 cm

These are complete test snapshot results. See the evaluation protocol for metric definitions and coverage. The current package changes only the embedded model name from the evaluated package; all 477 parameter tensors are identical.

Attribution

F2G-Pose integration code uses Apache-2.0. Third-party components and the Omni6DPose dataset retain their own terms. See THIRD_PARTY.md and the source repository.