Submission: SingleMultiPathEncoder-CVPR20/ITODD

Submission name
Submission time (UTC) Aug. 19, 2020, 5:33 p.m.
User MartinSmeyer
Dataset ITODD
Training model type Default
Training image type Synthetic (custom)
Description
Evaluation scores
bop19_average_recall:0.067
bop19_average_recall_mspd:0.140
bop19_average_recall_mssd:0.031
bop19_average_recall_vsd:0.031
bop19_average_time_per_image:0.177

Method: SingleMultiPathEncoder-CVPR20

User MartinSmeyer
Publication https://openaccess.thecvf.com/content_CVPR_2020/html/Sundermeyer_Multi-Path_Learning_for_Object_Pose_Estimation_Across_Domains_CVPR_2020_paper.html
Implementation https://github.com/DLR-RM/AugmentedAutoencoder/tree/multipath
Training image modalities RGB
Test image modalities RGB
Description

A single MultiPath Encoder model to predict the poses of all objects of all datasets in the BOP challenge (instead of training one model per object/dataset). Single object OpenGL renderings are used for training.

The single MultiPath Encoder achieves results similar to 108 separately trained AAE models used in the BOP19 challenge on pure RGB data. This result should serve as a baseline for learning-based pose-sensitive feature extraction that can scale to a large number of trained & untrained objects.

For instance segmentation we use dataset-wise trained MaskRCNNs (trained on BlenderProc PBR and real data).

Computer specifications Intel(R) Xeon(R) W-2133 CPU @ 3.60GHz + Intel Nvidia 1080 Ti