This repo is the official project repository of the paper 'Enhancing point cloud analysis via neighbor aggregation correction based on cross-stage structure correlation' and is mainly used for updating instructions, sharing experiment records (containing model weight), and handling issues. This project is based on OpenPoint framework, pleas refer to this online file to how to install and use. Paper is here.
The PDSA module we proposed is an improved version based on SAmodule, and can be used to replace the Set Abstraction module in existing works such as PointNext and PointVector to correct the feature distribution to achieve better performance.
Please follow the usage instructions of the openpoint code framework and place the config and model files in the corresponding paths. The code module of PDSA is in PPV2.py. Please replace the SA module in PointNext, Pointvector or other models accordingly.
The chechk point files in S3DIS, ScanObecjtNN and ModelNet are in huggingface, please download to use.
All the experimental Settings are in the corresponding yaml files under the cfg folder. Our experiment environment is Nvidia Geforce RTX 4090 24-GB GPU with a 12-core Intel Xeon(R) Platinum 8352V CPU @2.10Ghz.
@misc{shi2025enhancingpointcloudanalysis, title={Enhancing point cloud analysis via neighbor aggregation correction based on cross-stage structure correlation}, author={Jiaqi Shi and Jin Xiao and Xiaoguang Hu and Boyang Song and Hao Jiang and Tianyou Chen and Baochang Zhang}, year={2025}, eprint={2506.15160}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2506.15160}, }