Semi-Weakly Supervised Object Kinematic Motion Prediction (2023 CVPR)

Introduction

In this paper, we tackle the task of object kinematic motion prediction problem in a semi-weakly supervised manner. Our key observations are two-fold. First, although 3D dataset with fully annotated motion labels is limited, there are existing datasets and methods for object part semantic segmentation at large scale. Second, semantic part segmentation and mobile part segmentation is not always consistent but it is possible to detect the mobile parts from the underlying 3D structure. Towards this end, we propose a graph neural network to learn the map between hierarchical part-level segmentation and mobile parts parameters, which are further refined based on geometric alignment. This network can be first trained on PartNet-Mobility dataset with fully labeled mobility information and then applied on PartNet dataset with fine-grained and hierarchical part-level segmentation. The network predictions yield a large scale of 3D objects with pseudo labeled mobility information and can further be used for weakly-supervised learning with pre-existing segmentation. Our experiments show there are significant performance boosts with the augmented data for previous method designed for kinematic motion prediction on 3D partial scans.

Dataset

In our experiments, we use the PartNet-Mobility dataset and the PartNet dataset, please download these two dataset first. Ensure that the directory of PartNet-Mobility dataset is as follows:

partnet_mobility_root/
    └── dataset/ 
        ├── file_1/ 
                ├── images/
                ├── parts_render/
                ├── parts_render_after_merging/
                ├── point_sample/ 
                ├── textured_objs/
                ├── bounding_box.json
                ├── meta.json
                ├── mobility.urdf
                ├── mobility_v2.json
                ├── result.json
                ├── result_after_merging.json
                ├── semantics.txt
                ├── tree_hier.html
                └── tree_hier_after_merging.html
        ├── file_2
        ├── ... 
        └── file_n 

Ensure that the directory of PartNet dataset is as follows:

partnet_root/
    └── dataset/ 
        ├── Bottle/ 
                └── data_v0/
                    ├── file_1/
                            ├── objs/
                            ├── parts_render/
                            ├── parts_render_after_merging/
                            ├── point_sample/ 
                            ├── meta.json 
                            ├── result.json 
                            ├── result_after_merging.json
                            ├── tree_hier.html
                            └── tree_hier_after_merging.html 
                    ├── file_2/
                    ├── ...
                    └── file_n/
        ├── Chair
        ├── ... 
        └── TrashCan

Installation

We use two virtual environments in our project. The configuration of the first virtual environment are:

• Python 3.8
• cuda 11.2
• Pytorch 1.8.1
• Matlab 9.11.0.1837725 (R2021b)

This virtual environment is used to process the Partnet-Mobility dataset and the PartNet dataset, and train the models proposed in Section 3.2 and Section 3.3 of our papaer. To install this virtual environment:

conda create –n motion python=3.8
source activate motion
pip install –r requirements.txt

To facilitate the call of matlab functions in python, we need to install the matlab engine for python. Find the matlab installation path MATLAB_PATH, and then run the following command.

cd MATLAB_PATH/extern/engines/python && python setup.py install

The second virtual environment is to verify the effectiveness of the pseudo labels annotated by our method on the PartNet dataset. We use ANCSH to verify. Pelease refer to their repository to install the second virtual environment.

Usage

Before running the code, it is necessary to modify the config.json file and set the paths to the data for PartNet-Mobility and PartNet. The code running steps and descriptions of each folder are as follows:

1. TableComplete: Since the objects of the Table category in the PartNet-Mobility and the PartNet dataset miss internal structure, therefore we first complete the internal structure of the table.
2. NetworkDataExtract: The code of this folder extracts the network input data on the PartNet-Mobility and the PartNet dataset.
3. GnnMotion: Train the graph neural network proposed in Section 3.2 of our paper.
4. AxisSelection: Train the axis selection network proposed in Section 3.3 of our paper.
5. Predict: Use the graph neural network and the axis selection network to predict motion parameters on PartNet dataset, and generate mobility_v2.json file.
6. Extern: Convert the annotated PartNet data and boost the performance of ANCSH.

License

Our code is released under MIT License. See LICENSE file for details.

Citation

Please cite the paper in your publications if it helps your research:

@inproceedings{liu2023semi,
  title={Semi-Weakly Supervised Object Kinematic Motion Prediction},
  author={Liu, Gengxin and Sun, Qian and Huang, Haibin and Ma, Chongyang and Guo, Yulan and Yi, Li and Huang, Hui and Hu, Ruizhen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}

Acknowledgments

Our graph neural network and axis selection network are based on StructureNet and DGCNN, and the partial scans motion prediction network is based on ANCSH. Thanks these teams for their awesome codebase. Also thanks the team of DSG-Net for providing the generated shapes.