This repo contains the code for OOD-CV Workshop SSB Challenge 2024 (Open-Set Recognition Track)

Setup

Please refer SSB to setup,including: setting up a Kaggle account; setting up an SSB JSON config; installing python requirements; and installing this SSB Python package.

Download datasets

After setup, please refer the download part in SSB. This is an example:

from SSB.download import download_datasets
download_datasets(['cub', 'aircraft', 'scars', 'imagenet_1k', 'imagenet_21k'])

Get Open-Set Recognition (OSR) datasets

Only ImageNet is currently supported in the OSR challenge. The SSB split (i.e 'Easy' or 'Hard') should be specified in osr_split.

Documentation is given inside the function in SSB/get_datasets/get_osr_datasets_funcs.py.

Specifying eval_only=True means only the test datasets will be returned (the ImageNet-1k training set is not loaded). This is faster.

from SSB import get_osr_datasets
all_datasets = get_osr_datasets(dataset_name='imagenet',
                                osr_split='Hard', 
                                train_transform=None, 
                                test_transform=test_transform,
                                eval_only=True)

Our solution

Our approach is improved based on BASELINE, including the following two points:

• Multi-model joint inference for pre-trained weights of the same image size
• Multi-model joint inference based on data augmentation

Usage

If you want to use joint inference only, you can

python OOD-OSR/evaluate_osr_model_fusion.py

If you want to add data augment in joint inference, you can

python OOD-OSR/evaluate_osr_TTA.py

Citation

@InProceedings{vaze2022openset,
      title={Open-Set Recognition: a Good Closed-Set Classifier is All You Need?},
      author={Sagar Vaze and Kai Han and Andrea Vedaldi and Andrew Zisserman},
      booktitle={International Conference on Learning Representations},
      year={2022}
      }