Skip to content

Latest commit

 

History

History
70 lines (53 loc) · 2.11 KB

README.md

File metadata and controls

70 lines (53 loc) · 2.11 KB
Raw

[CVPR2022 Oral] FisherMatch: Semi-Supervised Rotation Regression via Entropy-based Filtering

Project page: https://yd-yin.github.io/FisherMatch/

Setup

Dependencies

  • PyTorch=1.8.1 + torchvision=0.9.1 + cudatoolkit=10.2
  • PyTorch3D (Installation via conda or pypi may be problematic. Build from source if necessary.)
  • Other dependencies 
    pip install opencv-python tqdm matplotlib scipy tensorboard configargparse lmdb pyyaml

Dataset

ModelNet10-SO(3)

Download ModelNet10-SO(3) dataset here and link it to data/ModelNet10-SO3

unzip ModelNet10-SO3.zip
ln -s $PWD/ModelNet10-SO3 $PROJECT_PATH/data/ModelNet10-SO3

Pascal3D+

Download Pascal3D+ (release1.1) dataset here and link it to data/pascal3d

unzip PASCAL3D+_release1.1.zip
ln -s $PWD/PASCAL3D+_release1.1 $PROJECT_PATH/data/pascal3d

Generate data annotations

python dataset_pascal.py

Usage

The default experiment setting is stored in settings/ssl.yml. To change the settings, you may either specify arguments in the command-line, or create settings/EXAMPLE.yml and pass it to --config argument.

If a value is specified in more than one way then: command line > config file values > defaults.

Train

python train.py [--config=settings/EXAMPLE.yml]

For experiments on Pascal3D+ dataset, we use different confidence thresholds for different categories and #labeled images. Please finetune the confidence threshold hyper-parameter to get better results.

The training process is logged by tensorboard.

Evaluation

python eval.py CKPT_FILES [--config=settings/EXAMPLE.yml]

Bibtex

@InProceedings{yin2022fishermatch,
  author={Yin, Yingda and Cai, Yingcheng and Wang, He and Chen, Baoquan},
  title={FisherMatch: Semi-Supervised Rotation Regression via Entropy-Based Filtering},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month={June},
  year={2022},
  pages={11164-11173}
}