Submission: Pix2Pose-BOP19-PBR/LM-O

Submission name
Submission time (UTC) June 12, 2020, 11:56 a.m.
User kirumang
Dataset LM-O
Training model type Default
Training image type Synthetic (only PBR images provided for BOP Challenge 2020 were used)
Description Conclusion: Using PBR images for training (instead of given synthetic images) dramatically improves the performance for our method (BOP score: 0.077 -> 0.281) . I highly recommend others to use PBR images for their training!
Evaluation scores
bop19_average_recall:0.281
bop19_average_recall_mspd:0.478
bop19_average_recall_mssd:0.210
bop19_average_recall_vsd:0.156
bop19_average_time_per_image:1.157

Method: Pix2Pose-BOP19-PBR

User kirumang
Publication Pix2Pose: Pixel-Wise Coordinate Regression of Objects for 6D Pose Estimation, ICCV 2019
Implementation https://github.com/kirumang/Pix2Pose
Training image modalities RGB
Test image modalities RGB
Description The results are obtained using the same code used for BOP'19 with the original paper backbone (trained from scratch, no pre-trained weights are used for Pix2Pose, for Mask-RCNN : MScoco pre-trained weights are used) This is to directly compare the effect of PBR images while using the exactly the same parameters for training and evaluations that we used for BOP'19 challenge. (In the BOP'19, we used synthetic images that are provided in the challenge without additional renderings)
Computer specifications intel i7-6700K, NVIDA GTX 1080, 16GB RAM