Check out our paper here which has been accepted to IEEE Secure and Trustworthy Machine Learning (SaTML) 2025!
In this work, we assess the effect of texture bias on real data decisions by exploring two key components:
-
Quantifying the Relationship Between Textures and Objects
We first measure the degree to which a model associates textures with object classes. This is done using texture data to generate Texture-Object Association Values (TAV). The figure below shows the strength of the top 50 associations for a ResNet-50 model trained on ImageNet: -
Identifying Textures Present in Object Images
Using the learned associations, we identify which textures are present in real-world images. This allows us to understand how texture influences model predictions on standard benchmarks. The following figure shows images from the ImageNet validation set that are identified as having a grid texture.
Please follow the instructions in datasets/README.md to download and prepare the datasets.
To compute the TAV using your model of choice (e.g., resnet50), run:
python3 main.py --model_name {MODEL_NAME_HERE} --dataset ptdThis will store the TAV results in the results/ directory.
Once the TAV has been computed, you can run TID on any dataset using:
python3 main.py --model_name {MODEL_NAME_HERE} --dataset {DATASET_NAME_HERE}The currently supported datasets include:
imagenet-valimagenet-a
If you use this work in your research, please cite the following paper:
@article{hoak2024err,
title={Err on the Side of Texture: Texture Bias on Real Data},
author={Hoak, Blaine and Sheatsley, Ryan and McDaniel, Patrick},
journal={arXiv preprint arXiv:2412.10597},
year={2024}
}