Submission: CosyPose-ECCV20-PBR-1VIEW/TUD-L

Download submission
Submission name
Submission time (UTC) May 4, 2022, 3:56 p.m.
User yann_labbe
Task Model-based 2D segmentation of seen objects
Dataset TUD-L
Training model type Default
Training image type Synthetic (only PBR images provided for BOP Challenge 2020 were used)
Description
Evaluation scores
AP:0.306
AP50:0.783
AP75:0.173
AP_large:0.110
AP_medium:0.333
AP_small:0.240
AR1:0.346
AR10:0.352
AR100:0.352
AR_large:0.149
AR_medium:0.366
AR_small:0.250
average_time_per_image:0.044

Method: CosyPose-ECCV20-PBR-1VIEW

User yann_labbe
Publication Labbé et al, CosyPose: Consistent multi-view multi-object 6D pose estimation, ECCV 2020
Implementation https://github.com/ylabbe/cosypose
Training image modalities RGB
Test image modalities RGB
Description

The method is the single-view (1 view) object pose estimation introduced in Labbé et al, CosyPose: Consistent multi-view multi-object 6D pose estimation, ECCV 2020.

For each dataset, we train three networks: a MaskRCNN detector (only 2D detections are used at test time), a model for coarse pose estimation, and a model for iterative refinement. The refinement network is ran for 4 iterations at test time.

Only provided PBR synthetic images are used on each dataset. We add data augmentation to the synthetic images as described in the paper. Pose networks are trained from scratch on all objects. MaskRCNN has a resnet50 FPN backbone pre-trained on COCO.

The timing includes detection and pose estimation on all detections. We do not use the targets file to filter out the detections that are not evaluated.

An extended version of the CosyPose paper with BOP challenge results is available here: https://drive.google.com/file/d/1YVemqAEb_aL174XPYD-wag-DW7mUnvzC/view The detection method is described in Sec 4.3.

Computer specifications CPU: 20-core Intel Xeon 6164 @ 3.2 GHz, GPU: Nvidia V100