Submission: EPOS-BOP20-PBR/T-LESS

Download submission
Submission name
Submission time (UTC) Aug. 19, 2020, 7:35 p.m.
User thodan
Task 6D localization of seen objects
Dataset T-LESS
Training model type Reconstucted
Training image type Synthetic (only PBR images provided for BOP Challenge 2020 were used)
Description
Evaluation scores
AR:0.467
AR_MSPD:0.619
AR_MSSD:0.403
AR_VSD:0.380
average_time_per_image:1.992

Method: EPOS-BOP20-PBR

User thodan
Publication Hodan, Barath, Matas, EPOS: Estimating 6D Pose of Objects with Symmetries, CVPR 2020
Implementation https://github.com/thodan/epos
Training image modalities RGB
Test image modalities RGB
Description

Only the synthetic "PBR-BlenderProc4BOP" RGB images provided for the BOP Challenge 2020 were used for training.

The results were achieved without any post-refinement of the estimated poses (i.e. without ICP, DeepIM, etc.).

Up to 5000 2D-3D correspondences with the highest confidence were used per image. Other hyper-parameters were set as described in the CVPR 2020 paper.

Computer specifications 14-core Intel Xeon E5-2680 v4 CPU, 252GB RAM, Nvidia P100 GPU