Submission: leaping from 2D to 6D/LM-O/lmo_0820

Download submission
Submission name lmo_0820
Submission time (UTC) Aug. 19, 2020, 3:26 p.m.
User jinhuiliu
Task 6D localization of seen objects
Dataset LM-O
Training model type Default
Training image type Synthetic (only PBR images provided for BOP Challenge 2020 were used)
Evaluation scores

Method: leaping from 2D to 6D

User jinhuiliu
Training image modalities RGB
Test image modalities RGB

We propose a 6DoF object pose estimation framework. Directly regressing object poses in 6d dimension is difficult due to the large search space. Instead, we employ a 2D object detector to firstly localize objects in the 2D image plane, and then predict poses within local regions. Moreover, the detected bounding boxes can be used to filter out some inaccurate object segmentation or reduce the regions to be segmented. Thus, the bounding boxes can either facilitate object keypoint localization or reduce computational cost of following pose estimation network since poses are now estimated in local regions. We train one network for each one object.

Computer specifications CPU: Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz, GPU: Tesla P40