Point MixSwap

This is the official repository of our ECCV 2022 paper, "Point MixSwap: Attentional Point Cloud Mixing via Swapping Matched Structural Divisions".

Dataset

Prepare ModelNet40 dataset by downloading it via this link. Then, extract the zip file and move to folder such that the structure becomes: data/modelnet40_ply_hdf5_2048.

Train

To train the model, DGCNN with ModelNet40, run the script below (e.g., using the first GPU device): CUDA_VISIBLE_DEVICES=0 python main.py --config=configs/config.yaml

Note: the code is tested in NVIDIA GeForce RTX 3090, using pyhton 3.9 with Ubuntu 18.04.5 LTS.

Config

Yaml config file can be found in configs folder. The file contains all the hyperparamters setup and other related configurations. Most of them are self-explanatory by looking the variable names. Here some details:

• MIXUP_LEVEL: where to perform the mixup data, which can be input, feature and both
  • input: perform mixup in the input model
  • feature: perform mixup in the feature model
  • both: perform mixup in both input and feature model

Some Visualizations

Cite

@article{pointmixswap,
          author = {Umam, Ardian and Yang, Cheng-Kun and Chuang, Yung-Yu and Chuang, Jen-Hui and Lin, Yen-Yu},
          title = {Point MixSwap: Attentional Point Cloud Mixing via Swapping Matched Structural Divisions},
          booktitle={European Conference on Computer Vision},
          year={2022}
}