Submission: ZebraPoseSAT-EffnetB4 + ICP (DefaultDetection)/TUD-L

Download submission
Submission name
Submission time (UTC) Oct. 15, 2022, 6:02 p.m.
User zebrapose
Task 6D localization of seen objects
Dataset TUD-L
Training model type Default
Training image type Synthetic + real
Description
Evaluation scores
AR:0.948
AR_MSPD:0.981
AR_MSSD:0.978
AR_VSD:0.885
average_time_per_image:0.500

Method: ZebraPoseSAT-EffnetB4 + ICP (DefaultDetection)

User zebrapose
Publication ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation, CVPR2022
Implementation https://github.com/suyz526/ZebraPose
Training image modalities RGB
Test image modalities RGB-D
Description

Based on the paper "ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation", CVPR 2022.

  • Training images: PBR image + Real images (if available)

  • Setting: One network per object was trained

  • 2D Bounding box: default detections ("synt+real" version) provided by the BOP organizers.

  • Modifications to the original ZebraPose paper:

    1. Added Symmetry-Aware Training (SAT). The network and loss functions are not changed. There will be a new ground truth for the sym. objects, details can be found in the Github Repository. Special thanks to Yongliang Lin for his contribution.

    2. We replace the Resnet34 backbone with EffnetB4 in ZebraPose. (Only replace the backbone in the pose estimation part)

    3. We apply ICP algorithm for pose refinement with depth image. The code for ICP is adapted from Cosypose.

  • The reported inference time included 2D detection time.

List of contributors:

  • German Research Center for Artificial Intelligence (DFKI), Augmented Vision department:

Yongzhi Su, Praveen Nathan, Torben Fetzer, Jason Rambach, Didier Stricker

  • Technical University Munich (TUM), CAMPAR:

Mahdi Saleh, Yan Di, Nassir Navab, Benjamin Busam, Federico Tombari

  • Zhejiang University (ZJU):

Yongliang Lin, Yu Zhang

Computer specifications Intel(R) Xeon(R) E-2146G CPU @ 3.50GHz, Nvidia RTX2080Ti