megayeye / clean-pvnet

Code for "PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation" CVPR 2019 oral

PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation

PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation
Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, Hujun Bao
CVPR 2019 oral
Project Page

Any questions or discussions are welcomed!

Introduction

Thanks Haotong Lin for providing the clean version of PVNet and reproducing the results.

The structure of this project is described in project_structure.md.

Installation

1. Set up the python environment:

   conda create -n pvnet python=3.7
   conda activate pvnet
   
   # install torch 1.1 built from cuda 9.0
   pip install torch==1.1.0 -f https://download.pytorch.org/whl/cu90/stable
   
   pip install Cython==0.28.2
   sudo apt-get install libglfw3-dev libglfw3
   pip install -r requirements.txt
   

2. Compile cuda extensions under lib/csrc:

   ROOT=/path/to/clean-pvnet
   cd $ROOT/lib/csrc
   export CUDA_HOME="/usr/local/cuda-9.0"
   cd dcn_v2
   python setup.py build_ext --inplace
   cd ../ransac_voting
   python setup.py build_ext --inplace
   cd ../nn
   python setup.py build_ext --inplace
   cd ../fps
   python setup.py
   

3. Set up datasets:

   ROOT=/path/to/clean-pvnet
   cd $ROOT/data
   ln -s /path/to/linemod linemod
   ln -s /path/to/linemod_orig linemod_orig
   ln -s /path/to/occlusion_linemod occlusion_linemod
   
   # the following is used for tless
   ln -s /path/to/tless tless
   ln -s /path/to/cache cache
   ln -s /path/to/SUN2012pascalformat sun
   

Download datasets which are formatted for this project:

1. linemod
2. linemod_orig: The dataset includes the depth for each image.
3. occlusion linemod
4. truncation linemod: Check TRUNCATION_LINEMOD.md for the information about the Truncation LINEMOD dataset.
5. Tless: cat tlessa* | tar xvf - -C ..
6. Tless cache data: It is used for training and testing on Tless.
7. SUN2012pascalformat

Testing

Testing on Linemod

We provide the pretrained models of objects on Linemod, which can be found at here.

Take the testing on cat as an example.

1. Prepare the data related to cat:

   python run.py --type linemod cls_type cat
   

2. Download the pretrained model of cat and put it to $ROOT/data/model/pvnet/cat/199.pth.
3. Test:

python run.py --type evaluate --cfg_file configs/linemod.yaml model cat
python run.py --type evaluate --cfg_file configs/linemod.yaml test.dataset LinemodOccTest model cat

4. Test with icp:

python run.py --type evaluate --cfg_file configs/linemod.yaml model cat test.icp True
python run.py --type evaluate --cfg_file configs/linemod.yaml test.dataset LinemodOccTest model cat test.icp True

Testing on Tless

We provide the pretrained models of objects on Tless, which can be found at here.

1. Download the pretrained models and put them to $ROOT/data/model/pvnet/.
2. Test:

   python run.py --type evaluate --cfg_file configs/tless/tless_01.yaml
   

Visualization

Take the cat as an example.

1. Prepare the data related to cat:

   python run.py --type linemod cls_type cat
   

2. Download the pretrained model of cat and put it to $ROOT/data/model/pvnet/cat/199.pth.
3. Visualize:

python run.py --type visualize --cfg_file configs/linemod.yaml model cat

If setup correctly, the output will look like

Training

Training on Linemod

1. Prepare the data related to cat:

   python run.py --type linemod cls_type cat
   

2. Train:

   python train_net.py --cfg_file configs/linemod.yaml model mycat cls_type cat
   

The training parameters can be found in project_structure.md.

Training on Tless

Train:

python train_net.py --cfg_file configs/tless/tless_01.yaml

Tensorboard

tensorboard --logdir data/record/pvnet

If setup correctly, the output will look like

Training on the custom object

1. Create a dataset using https://github.com/F2Wang/ObjectDatasetTools
2. Organize the dataset as the following structure:

   ├── /path/to/dataset
   │   ├── model.ply
   │   ├── camera.txt
   │   ├── rgb/
   │   │   ├── 0.jpg
   │   │   ├── ...
   │   │   ├── 1234.jpg
   │   │   ├── ...
   │   ├── mask/
   │   │   ├── 0.png
   │   │   ├── ...
   │   │   ├── 1234.png
   │   │   ├── ...
   │   ├── pose/
   │   │   ├── 0.npy
   │   │   ├── ...
   │   │   ├── 1234.npy
   │   │   ├── ...
   │   │   └──
   

3. Create a soft link pointing to the dataset:

   ln -s /path/to/custom_dataset data/custom
   

4. Process the dataset:

   python run.py --type custom
   

5. Train:

   python train_net.py --cfg_file configs/linemod.yaml train.dataset CustomTrain test.dataset CustomTrain model mycat train.batch_size 4
   

6. Watch the training curve:

   tensorboard --logdir data/record/pvnet
   

7. Visualize:

   python run.py --type visualize --cfg_file configs/linemod.yaml train.dataset CustomTrain test.dataset CustomTrain model mycat
   

8. Test:

   python run.py --type evaluate --cfg_file configs/linemod.yaml train.dataset CustomTrain test.dataset CustomTrain model mycat
   

An example dataset can be downloaded at here.

Citation

If you find this code useful for your research, please use the following BibTeX entry.

@inproceedings{peng2019pvnet,
  title={PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation},
  author={Peng, Sida and Liu, Yuan and Huang, Qixing and Zhou, Xiaowei and Bao, Hujun},
  booktitle={CVPR},
  year={2019}
}