[WACV'23] PIDS: Joint Interaction-Dimension Search for 3D Point Cloud

This is the official implementation of our WACV'23 paper, "PIDS: Joint Interaction-Dimension Search for 3D Point Cloud" [https://openaccess.thecvf.com/content/WACV2023/html/Zhang_PIDS_Joint_Point_Interaction-Dimension_Search_for_3D_Point_Cloud_WACV_2023_paper.html]. The framework is based developed upon PyTorch.

Highlights

• Point-operators that compose both point interactions and point dimensions are implemented under pids_core/models/blocks.py. The first-order point interaction is implemented under pids_core/models/attention.py.

• Dense-Sparse predictor that intuited by the idea of "Wide & Deep Learning" to learn a better architecture prediction. The predictor is called in pids/predictor/predictor_model_zoo.py and implemented in nasflow/algo/surrogate/predictor.

Starter Guide

• Installing prerequisite. We provide a conda environment file pids_env.yaml for you to start with. To install all files that fullfill the requirement, you can run the following command:

conda create -n pids -f pids_env.yaml

• To get the dataset, please refer to tutorials/architecture_search_guidelines.ipynb and place it under the folder of your interest.

• (Recommended). It's better to re-compile the cpp wrappers to fit the runtime of your current machine. Simply go to pids_core/cpp_wrappers and run compile_wrappers.sh gets the job done.

Architecture Search Process

Architecture search code is released for easy use and development. Please refer to tutorials/architecture_search_guidelines.ipynb for detailed instructions and hyperparameter guidance.

Training the searched architectures from scratch

The searched architectures are revealed in pids_search_space/arch_genotype.py, defined as a composition of block args representing architecture compositions. If you run the search by yourself, you can modify pids_search_space/arch_genotype.py to incorporate the changes that you have. Please refer to final_evaluation.ipynb for detailed training scripts and hyperparameter definitions.

Obtaining validation/test accuracy.

The evaluation process on 3D-point cloud is not as straight-forward as image classification. To obtain the validation/test accuracy, a voting mechansim is required that runs the point cloud multiple times and average the predictions.

Evaluation

Our paper presents the following results. More detailed comparisons can be seen on the original paper.

Please refer to testing/val_models.py for the testing process. The general usage should be:

python testing/val_models.py --result_path [CKPT_PATH]

Results

SemanticKITTI (08-val)

Method	Parameter (M)	Multiply-Accumulates (G)	Latency (ms)	mIOU (%)	Notes
KPConv	14.8	60.9	221 (164 + 57)	59.2
PIDS (Second-order)	0.97	4.7	160 (103 + 57)	60.1
PIDS (NAS)	0.57	4.4	169 (112 + 57)	62.4	Checkpoint
PIDS (NAS, 2x)	1.36	11.0	206 (149 + 57)	64.1	Checkpoint

S3DIS (Area-05)

Refer to training/train_S3DIS.py for implementation details.

Method	mIOU	ceil.	floor	wall	beam	col.	wind.	door	chair	table	book.	sofa	board	clut.
KPConv	65.4	92.6	97.3	81.4	0.0	16.5	54.5	69.5	90.1	80.2	74.6	66.4	63.7	58.1
PIDS	67.2	93.6	98.3	81.6	0.0	32.2	51.5	73.2	90.7	82.5	73.3	64.7	71.6	60.0

Checkpoint is available here.

ModelNet40

Method	Parameter (M)	Overall Accuracy (%)	Notes
KPConv	14.9	92.9
PIDS (Second-order)	1.25	92.6
PIDS (NAS)	0.56	93.1	Checkpoint
PIDS (NAS, 2x)	1.21	93.4	Checkpoint

Acknowledgement

This project is in part supported by the following grants: NSF-2112562, NSF-1937435, and ARO W911NF-19-2-0107, and CAREER-2048044. We also acknowledge the original KPConv-PyTorch project [https://github.com/HuguesTHOMAS/KPConv-PyTorch] to provide backbone implementations for our work.

Citation

If you would like to use this repository to develop your research work, feel free to use the citation below:

@inproceedings{zhang2023pids,
  title={PIDS: Joint Point Interaction-Dimension Search for 3D Point Cloud},
  author={Zhang, Tunhou and Ma, Mingyuan and Yan, Feng and Li, Hai and Chen, Yiran},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={1298--1307},
  year={2023}
}