Method: FoundPose+FeatRef+Megapose-5hyp

User epi
Publication
Implementation
Training image modalities RGB
Test image modalities RGB
Description

The presented results were achieved by FoundPose with the featuremetric refinement and additional MegaPose [D] refinement (row 12 of Table 1 in [A]).

In this submission, FoundPose uses default CNOS-FastSAM [B] segmentations provided for BOP'23. For pose estimation, the method uses features from layer 18 of DINOv2 (ViT-L) with registers [C]. For each object instance, 5 coarse pose hypotheses are estimated and refined. The pose with the highest MegaPose score is selected as the final pose estimate.

Note that FoundPose doesn't do any task-specific training -- it only uses frozen FastSAM (via CNOS) and frozen DINOv2. The only component that is used in this submission and trained in a task-specific manner is the MegaPose refiner (we used weights from the official MegaPose repository and didn't train them further).

[A] Anonymous: FoundPose: Unseen Object Pose Estimation with Foundation Features.

[B] Nguyen et al.: CNOS: A Strong Baseline for CAD-based Novel Object Segmentation, ICCVW 2023.

[C] Darcet et al.: Vision transformers need registers, arXiv 2023.

[D] Labbé et al.: MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare, CoRL 2022.

Computer specifications Tesla P100 16GB

Public submissions

Date Submission name Dataset
2023-11-16 18:13 FoundPose+FeatRef+Megapose-5hyp ITODD
2023-11-16 19:20 FoundPose+FeatRef+Megapose-5hyp HB
2024-01-26 14:59 FoundPose+FeatRef+Megapose-5hyp TUD-L
2024-01-26 15:01 FoundPose+FeatRef+Megapose-5hyp IC-BIN
2024-01-27 10:43 FoundPose+FeatRef+Megapose-5hyp YCB-V
2024-01-27 10:43 FoundPose+FeatRef+Megapose-5hyp T-LESS
2024-01-27 10:43 FoundPose+FeatRef+Megapose-5hyp LM-O