RADAM: Texture Recognition through Randomized Aggregated Encoding of Deep Activation Maps

https://doi.org/10.1016/j.patcog.2023.109802
https://arxiv.org/abs/2303.04554

RADAM uses deep neural networks (CNNs) available in timm==0.6.7 for texture feature extraction, and then "classic" machine learning classification is done with scikit-learn classifiers. Several dataloaders are available for texture benchmarks, see datasets.py. RADAM works practically with any CNN architecture compatible with timm's features_only=True mode. See standalone_RADAM_example.py for getting started.

RADAM achieves state-of-the-art texture analysis performance without ever fine-tuning the pre-trained backbones, and using only a linear SVM for classification.

Usage of the RADAM module

• Check standalone_RADAM_example.py

net = timm.create_model(model, features_only=True, pretrained=True, output_stride=8)
texture_representation = RADAM(device, z, (w,h))(net(input_batch))

Requirements

• conda version : 4.10.3
• python version : 3.9.5.final.0
• And everything else inside requirements.yml

conda env create -f requirements.yml
conda activate timm_tfe

• You may need to add conda-forge to your anaconda channel list in order to find some of the package versions we used:

conda config --add channels conda-forge

• The code should work with newer versions of torch and timm, but may require minimal adjustments (especially for the timm models, you should check their exact name on your version)

Setup used for experiments

• Linux Ubuntu x86-64 18.04.5 LTS (Bionic Beaver)
• Intel(R) Core(TM) i7-7820X CPU @ 3.60GHz, 64GB RAM
• GTX 1080 ti, NVIDIA driver Version: 470.141.03, CUDA Version: 11.4

Additional resources

• Dataloaders for many texture datasets, see datasets.py
• Scripts for experimentation, see classify_features.py and others

Usage of the script for experiments

• See python classify_features.py --help for usage information. An example with RADAM:

python classify_features.py --model convnext_nano --pooling RAEspatial --M 4 --depth all --dataset DTD --data_path /datasets/ --output_path /results/

Pay attention to args:

• --data_path (path to load/download datasets)
• --output_path (path to save extracted features and classification results, need 2 subfolders inside: feature_matrix/ and classification/)

Credits

If you use our code or methods, please cite the paper:

Leonardo Scabini, Kallil M. Zielinski, Lucas C. Ribas, Wesley N. Gonçalves, Bernard De Baets, Odemir M. Bruno, RADAM: Texture recognition through randomized aggregated encoding of deep activation maps, Pattern Recognition, Volume 143, 2023, 109802, ISSN 0031-3203, https://doi.org/10.1016/j.patcog.2023.109802.

@article{scabini2023radam,
  title={RADAM: Texture Recognition through Randomized Aggregated Encoding of Deep Activation Maps},
  author={Scabini, Leonardo and Zielinski, Kallil M and Ribas, Lucas C and Gonalves, Wesley N and De Baets, Bernard and Bruno, Odemir M},
  journal={Pattern Recognition},
  pages={109802},
  year={2023},
  publisher={Elsevier}
}

---