UKPGAN: A General Self-Supervised Keypoint Detector

Yang You, Wenhai Liu, Yanjie Ze, Yong-Lu Li, Weiming Wang, Cewu Lu

CVPR 2022

UKPGAN is a self-supervised 3D keypoint detector on both rigid/non-rigid objects and real scenes. Note that our keypoint detector solely depends on local features and is both translational and rotational invariant.

News

• [2022.05] Pytorch version is now supported! Try it in the branch pytorch!
• [2022.05] SHOT descriptor is now supported with Pytorch, and it is much faster than the original PerfectMatch desciptor!

Change Logs

• [2022.05.24] Upload Pytorch version of code and SHOT descriptors.

Contents

• Overview
• Installation
• Train on ShapeNet Models
• Test on ShapeNet Models
• Train on SMPL Models
• Test on SMPL Models
• Test on Real-world Scenes
• Pretrained Models
• Pytorch Version
• Related Projects
• Citation

Overview

This repo is a TensorFlow implementation of our work UKPGAN.

Installation

Create Conda Environments

conda env create -f environment.yml

Compile smoothed density value (SDV) source files

First install Pybind11 and PCL C++ dependencies. Then run the following command to build the SDV feature extractor:

cd sdv_src
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make
cd ../..

If you want to visualize keypoint results, you will need to install open3d.

Train on ShapeNet Models

Prepare Data

Download ShapeNet point clouds from KeypointNet and unzip pcds folder to the root.

Category Configuration

Change the category name cat_name to what you want in config/config.yaml.

Start Training

Open a separate terminal to monitor training process:

visdom -port 1080

Then run (e.g., chair):

python train.py cat_name=chair

Test on ShapeNet Models

Evaluate IoU

Once trained, to evaluate the IoU with human annotations, first download KeypointNet data (you may only download the category that you wish to evaluate), then run

python eval_iou.py --kpnet_root /your/kpnet/root

Visualization

To test and visualize on ShapeNet models, run:

python visualize.py --type shapenet --nms --nms_radius 0.1

Train on SMPL Models

Prepare Data

You should register on SMPL website and download the model. We follow this repo to pre-process the model to generate model.pkl. Place model.pkl into data/model.pkl.

Start Training

The following command start training on SMPL models on the fly:

python train.py cat_name=smpl symmetry_factor=0

Test on SMPL Models

Visualization

To test and visualize on SMPL models, run:

python visualize.py --type smpl --nms --nms_radius 0.2 --kp_num 10

Test on Real-world Scenes

For this task, we use the model that is trained on a large collection of ShapeNet models (across 10+ categories), called universal.

Prepare Data

You will need to download data from 3DMatch. We also provide a demo scene for visualization.

Visualization

To test and visualize on 3DMatch, run:

python visualize.py --type 3dmatch --nms --nms_radius 0.05

Pretrained Models

We provide pretrained models on Google Drive.

Pytorch Version

We have now implemented the Pytorch version of UKPGAN, and can be trained with SHOT descriptors, which are much faster than the PerfectMatch descriptors!

Training on ShapeNet

Directly clone the repo and run `git checkout pytorch` to switch to the pytorch version. Then run `python train.py`. You would get the following results when trained on ShapeNet chairs:

Related Projects

• KeypointNet: A Large-scale 3D Keypoint Dataset Aggregated from Numerous Human Annotations
• TopNet: Structural Point Cloud Decoder
• The Perfect Match: 3D Point Cloud Matching with Smoothed Densities

Citation

If you find our algorithm useful in your research, please consider citing:

@inproceedings{you2022ukpgan,
  title={UKPGAN: A General Self-Supervised Keypoint Detector},
  author={You, Yang and Liu, Wenhai and Ze, Yanjie and Li, Yong-Lu and Wang, Weiming and Lu, Cewu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}