README.md

metadata

task_categories:
  - image-classification
  - point-cloud-classification
  - point-cloud-recognition

Datasets

1. We conduct experiments on three new 3D domain generalization (3DDG) benchmarks proposed by us, as introduced in the next section.

   • base-to-new class generalization (base2new)
   • cross-dataset generalization (xset)
   • few-shot generalization (fewshot)

2. The structure of these benchmarks should be organized as follows.

    /path/to/Point-PRC
    |----data # placed in the same level of `trainers`, `weights`, etc. 
        |----base2new
            |----modelnet40
            |----scanobjectnn
            |----shapenetcorev2
        |----xset
            |----corruption
            |----dg
            |----sim2real
            |----pointda
        |----fewshot
            |----modelnet40
            |----scanobjectnn
            |----shapenetcorev2

3. You can find the usage instructions and download links of these new 3DDG benchmarks in the following section.

New 3DDG Benchmarks

Base-to-new Class Generalization

1. The datasets used in this benchmark can be downloaded according to the following links.

   • ModelNet40
   • S-OBJ_ONLY
   • S-OBJ_BG
   • S-PB_T50_RS
   • ShapeNetCoreV2

2. The following table shows the statistics of this benchmark.

   

Cross-dataset Generalization

1. The datasets used in this benchmark can be downloaded according to the following links.

   • OOD Generalization
     • OmniObject3d (Omin3D)
   • Data Corruption
     • ModelNet-C (7 types of corruptions)
       • add global outliers, add local outliers, dropout global structure, dropout local region, rotation, scaling, jittering
   • Sim-to-Real
   • PointDA

2. The following table shows the statistics of this benchmark.

   

Few-shot Generalization

1. Although this benchmark contains same datasets as the Base-to-new Class, it investigates the model generalization under extremely low-data regime (1, 2, 4, 8, and 16 shots), which is quite different from the evaluation setting in Base-to-new Class Generalization.

2. The following table shows the statistics of this benchmark.

   

Citation

1. If you find our paper and datasets are helpful for your project or research, please cite our work as follows.

  @inproceedings{sun24pointprc,
      title={Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis},
      author={Sun, Hongyu and Ke, Qiuhong and Wang, Yongcai and Chen, Wang and Yang, Kang and Li, Deying and Cai, Jianfei},
      booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS)},
      year={2024},
      url={https://openreview.net/forum?id=g7lYP11Erv}
  }