Submission: FFB6D-CVPR21-PBR-NoRefinement/LM-O

Submission name
Submission time (UTC) Dec. 27, 2021, 8:30 a.m.
User YishengHe
Dataset LM-O
Training model type Default
Training image type Synthetic (only PBR images provided for BOP Challenge 2020 were used)
Description FFB6D is a full flow bidirectional fusion network designed for 6D pose estimation from a single RGBD image. We only use the PBR data for training. The predicted result is not refined by any iterative refinement algorithms, i.e., ICP. Difference from the original FFB6D: - Use a single unified network for all object classes as all GT labels are provided in PBR. - Regenerate object SIFT-FPS 3D keypoint from BOP mesh models as their object coordinates are different from the original.
Evaluation scores
bop19_average_recall:0.687
bop19_average_recall_mspd:0.792
bop19_average_recall_mssd:0.718
bop19_average_recall_vsd:0.550
bop19_average_time_per_image:0.189

Method: FFB6D-CVPR21-PBR-NoRefinement

User YishengHe
Publication Yisheng He et al.: FFB6D: A Full Flow Bidirectional Fusion Network for 6D Pose Estimation, CVPR 2021 (Oral)
Implementation https://github.com/ethnhe/FFB6D
Training image modalities RGB-D
Test image modalities RGB-D
Description

FFB6D is a full flow bidirectional fusion network designed for 6D pose estimation from a single RGBD image. We only use the PBR data for training. The predicted result is not refined by any iterative refinement algorithms, i.e., ICP.

Computer specifications CPU: Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz; GPU: GeForce RTX 2080Ti