The official implementation of When Text and Images Don't Mix: Bias-Correcting Language-Image Similarity Scores for Anomaly Detection by Adam Goodge, Bryan Hooi and Wee Siong Ng and to appear in The 35th British Machine Vision Conference (BMVC2024).

The code is organized as follows:

Files

• clip : contains files for instantiating the CLIP model. Please download the model weights first.
• imagenet_utils : contains files and utility functions for ImageNet data.
• data_utils.py : contains functions for downloading and instantiating data.
• cifar_eval.py : the main file for cifar10 and cifar100 experiments.
• imagenet_eval.py : the main file for tinyimagenet experiments.

Citation

**@article{goodge2024text,
  title={When Text and Images Don't Mix: Bias-Correcting Language-Image Similarity Scores for Anomaly Detection},
  author={Goodge, Adam and Hooi, Bryan and Ng, Wee Siong},
  journal={arXiv preprint arXiv:2407.17083},
  year={2024}
}**