StablePose: Learning 6D Object Poses from Geometrically Stable Patches
CVPR 2021
Created by Yifei Shi, Junwen Huang, Xin Xu, Yifan Zhang and Kai Xu
This repository includes:
-
lib: the core Python library for networks and loss
** lib/loss_*dataset.py: symmetrynet loss caculation for respective dataset
** lib/network_*dataset.py: network architecture for the respective dataset -
datasets: the dataloader and training/testing lists
** datasets/tless/dataset.py: the training dataloader for tless dataset
** datasets/tless/dataset_test.py: the evaluation dataloader for tless dataset
** datasets/tless/dataset_config/*.txt: training and testing splits for tless dataset** datasets/linemod/dataset_lmo.py: the training dataloader for linemod dataset
** datasets/linemod/dataset_lmo_test.py: the evaluation dataloader for linemod dataset
** datasets/linemod/dataset_config/*.txt: training and testing splits for linemod dataset** datasets/nocs/dataset_nocs.py: the training dataloader for nocs dataset
** datasets/nocs/dataset_nocs_eval.py: the evaluation dataloader for nocs dataset
** datasets/nocs/dataset_config/*.txt: training and testing splits for nocs dataset
To train StablePose on T-LESS dataset, run
python train_tless.py
To train StablePose on NOCS-REAL275 dataset, run
python train_nocs.py
To train StablePose on LMO dataset, run
python train_lmo.py
To evaluate instace-level datasets: T-LESS and LMO, use the code here https://github.com/thodan/bop_toolkit.
To test StablePose on T-LESS dataset, run
python test_tless.py
To test/evaluate StablePose on LMO dataset, run
python test_lmo.py
The above scripts will create the required csv files in https://github.com/thodan/bop_toolkit.
To test/evaluate StablePose on NOCS-REAL275 dataset, run
python test_nocs.py
The detection results lmo can be found at this link.
The detection results tless can be found at this link.
The trained model for lmo can be found at this link.
The trained model for tless can be found at this link.
For full training and testing dataset: here (baidu yunpan, password: cqqx).
Note for simplicity and fair comparison, we combine all the categories in a single model for both instance-level and category-level in this repo, wich is different from the implementation in our paper.