Submission: G2L1/TUD-L/G2L_1_ICP

Download submission
Submission name G2L_1_ICP
Submission time (UTC) July 20, 2020, 11:03 p.m.
User FindTruth_19
Task Model-based 6D localization of seen objects
Dataset TUD-L
Training model type Default
Training image type Real
Description Our method is based on G2L-CVPR2020. We add an ICP component to further improve the performance. We train three different networks for three objects but with same parameters. For both 2D detector and G2L, we only use real training images. For the 2D detector, we use YOLOv3 as the backbone and train 200 epochs. For G2L we train 10 epochs.
Evaluation scores
AR:0.760
AR_MSPD:0.828
AR_MSSD:0.865
AR_VSD:0.585
average_time_per_image:0.225

Method: G2L1

User FindTruth_19
Publication Wei Chen et al: G2L-Net, CVPR 2020
Implementation
Training image modalities RGB-D
Test image modalities RGB-D
Description
Computer specifications