Submitted to: BOP Challenge 2023

Training data: real + provided PBR

Used 3D models: Reconstructed for T-LESS, default for other datasets

Notes:

Authors: Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji (Tsinghua University).

In the PBRReal-RGB-MModel setting, for LMO, HB, ICBIN and ITODD datasets, we only use the provided synthetic training data (PBR) in training. While for YCBV, TUDL, TLESS, we use the provided real data and synthetic data (PBR) in training. MModel means for each dataset, we trained a separate model for each object.

For detection, we adopted yolox as the detection method. Otherwise, stronger data augmentation and ranger optimizer has been used.

For pose estimation, the difference between our GDRNPP and the CVPR-version GDR-Net mainly includes:

• Domain Randomization: We used stronger domain randomization operations than the conference version during training.
• Network Architecture: We used a more powerful backbone Convnext rather than resnet-34, and two mask heads for predicting amodal mask and visible mask separately.
• Other training details, include learning rate, weight decay, visible threshold, and bounding box type.