Official Implementation for 'Unified Framework for Open-World Compositional Zero-shot Learning' in WACV 2025

Hirunima Jayasekara, Nirat Saini, Khoi Pham, Abhinav Shrivastava

Paper | Project Page

---

Setup

We provide an environment.yml file that can be used to create a Conda environment.

conda env create -f environment.yml
conda activate owczsl

Dataset

Pre-requisites:

Update the path for dataset images and log file in the config/*.yml files. Download the pre-trained models.

To download datasets,

sh download_data.sh

Training

To run the model for MIT-States Dataset:

python train.py with cfg=config/mit-states.yml per_gpu_batchsize=32 num_freeze_layers=0 lr_transformer=3.5e-6 lr=3.6e-6 lr_cross=1e-6 k=3 offset_val=0.1 neta=0.01

Evaluation

To evaluate the model for MIT-States Dataset:

python test.py  with cfg=config/mit-states.yml

Results

Open world performance on MIT-States, C-GQA and VAW-CZSL. As evaluation matrices we refer to AUC with seen and unseen compositions with different bias terms along with HM.

Citation

@INPROCEEDINGS{jayasekara2025unified,
  author={Jayasekara, Hirunima and Pham, Khoi and Saini, Nirat and Shrivastava, Abhinav},
  booktitle={2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)}, 
  title={Unified Framework for Open-World Compositional Zero-Shot Learning}, 
  year={2025},
  volume={},
  number={},
  pages={2706-2714}}