Gen6D

Gen6D is able to estimate 6DoF poses for unseen objects like the following video.

Project page | Paper

Todo List

• Pretrained models and evaluation codes.
• Pose estimation on custom objects.
• Training codes.

Installation

Required packages are list in requirements.txt. To determine how to install PyTorch along with CUDA, please refer to the pytorch-documentation

Download

1. Download pretrained models, GenMOP dataset and processed LINEMOD dataset at here.
2. Organize files like

Gen6D
|-- data
    |-- model
        |-- detector_pretrain
            |-- model_best.pth
        |-- selector_pretrain
            |-- model_best.pth
        |-- refiner_pretrain
            |-- model_best.pth
    |-- GenMOP
        |-- chair 
            ...
    |-- LINEMOD
        |-- cat 
            ...

Evaluation

# Evaluate on the object TFormer from the GenMOP dataset
python eval.py --cfg configs/gen6d_pretrain.yaml --object_name genmop/tformer

# Evaluate on the object cat from the LINEMOD dataset
python eval.py --cfg configs/gen6d_pretrain.yaml --object_name linemod/cat

Metrics about ADD-0.1d and Prj-5 will be printed on the screen.

Qualitative results

3D bounding boxes of estimated poses will be saved in data/vis_final/gen6d_pretrain/genmop/tformer. Ground-truth is drawn in green while prediction is drawn in blue.

Intermediate results about detection, viewpoint selection and pose refinement will be saved in data/vis_inter/gen6d_pretrain/genmop/tformer.

This image shows detection results.

This image shows viewpoint selection results. The first row shows the input image to the selector. The second row shows the input images rotated by the estimated in-plane rotation (left column) or the ground-truth in-plane rotation(right column) Subsequent 5 rows show the predicted (left) or ground-truth (right) 5 reference images with nearest viewpoints to the input image.

This image shows the pose refinement process. The red bbox represents the input pose, the green one represents the ground-truth and the blue one represents the output pose for the current refinement step.

Pose estimation on custom objects

Please refer to custom_object.md

Training

1. Download processed co3d data (co3d.tar.gz), google scanned objects data (google_scanned_objects.tar.gz) and ShapeNet renderings (shapenet.tar.gz) at here.
2. Download COCO 2017 training set.
3. Organize files like

Gen6D
|-- data
    |-- GenMOP
        |-- chair 
            ...
    |-- LINEMOD
        |-- cat 
            ...
    |-- shapenet
        |-- shapenet_cache
        |-- shapenet_render
        |-- shapenet_render_v1.pkl
    |-- co3d_256_512
        |-- apple
            ...
    |-- google_scanned_objects
        |-- 06K3jXvzqIM
            ...
    |-- coco
        |-- train2017

4. Train the detector

python train_model.py --cfg configs/detector/detector_train.yaml

5. Train the selector

python train_model.py --cfg configs/selector/selector_train.yaml

6. Prepare the validation data for training refiner

python prepare.py --action gen_val_set \
                  --estimator_cfg configs/gen6d_train.yaml \
                  --que_database linemod/cat \
                  --que_split linemod_val \
                  --ref_database linemod/cat \
                  --ref_split linemod_val

python prepare.py --action gen_val_set \
                  --estimator_cfg configs/gen6d_train.yaml \
                  --que_database genmop/tformer-test \
                  --que_split all \
                  --ref_database genmop/tformer-ref \
                  --ref_split all 

This command will generate the information in the data/val, which will be used in producing validation data for the refiner. 7. Train the refiner

python train_model.py --cfg configs/refiner/refiner_train.yaml

8. Evaluate all components together.

# Evaluate on the object TFormer from the GenMOP dataset
python eval.py --cfg configs/gen6d_train.yaml --object_name genmop/tformer

# Evaluate on the object cat from the LINEMOD dataset
python eval.py --cfg configs/gen6d_train.yaml --object_name linemod/cat

Acknowledgements

In this repository, we have used codes or datasets from the following repositories. We thank all the authors for sharing great codes or datasets.

• PVNet
• hloc
• COLMAP
• ShapeNet
• COCO
• Co3D
• Google Scanned Objects
• MVSNet_pl
• AnnotationTools

We provide a paper list about recent generalizable 6-DoF object pose estimators at https://github.com/liuyuan-pal/Awsome-generalizable-6D-object-pose.

Citation

@inproceedings{liu2022gen6d,
  title={Gen6D: Generalizable Model-Free 6-DoF Object Pose Estimation from RGB Images},
  author={Liu, Yuan and Wen, Yilin and Peng, Sida and Lin, Cheng and Long, Xiaoxiao and Komura, Taku and Wang, Wenping},
  booktitle={ECCV},
  year={2022}
}