Training and testing scripts for ResAD++ in the full-shot setting

[LFDarwin]

Could you provide the training and testing scripts for ResAD++ in the full-shot setting? Thank you very much

[xcyao00]

The scripts for training and testing are the same. In the paper, it is mentioned that under full shot setting, To avoid excessive reference features, we did not extract all normal sample features as references. We matched each sample with 10 spatially-aligned samples as references. I uploaded these aligned samples in the ``aligned'' directory. To use these samples, you can simply add the following code to the validate function:

normal_image_paths = get_normal_image_paths('/path/to/your/dataset', class_name, dataset='btad')  # normal train images
matched_indices = np.load(f'./aligned/indices/{dataset_name}/{class_name}_knn_indices.npy')

indices = matched_indices[idx]
rimage_paths = [normal_image_paths[ind] for ind in indices]
rimages = load_and_transform_vision_data(rimage_paths, device)
with torch.no_grad():
    ref_features = encoder.encode_image_from_tensors(rimages.to(device))
    for l in range(len(ref_features)):
        _, _, c = ref_features[l].shape
        ref_features[l] = ref_features[l].reshape(-1, c)

These code are already added in the validate function, you can just uncomment the corresponding code.

[xcyao00]

For obtaining the aligned reference samples, we follow the method in the paper: ``Target before shooting: Accurate anomaly detection and localization under one millisecond via cascade patch retrieval''. We provide a reference implementation in the file cpr.py.

[LFDarwin]

Okay, I will give it a try, and thank you again for your valuable contribution