Appreciation and Inquiry Regarding WinCLIP Code Implementation for Anomaly Detection

[JunseokLee3]

I wanted to extend my sincerest gratitude for the restored WinCLIP code you provided. Your code is impeccable and has been a tremendous learning resource for someone like myself who is not well-versed in coding.

However, I have a question regarding the application of red marks for the zero-shot and one-shot examples illustrated in the code. I utilized your approach to apply red marks as follows:

for key in scores:
    score = scores[key][idx]
    # Create a mask where the score is above a threshold, e.g., 0.5
    threshold_value = np.max(score) * 0.5 
    score_mask = score > threshold_value
    score_mask_img = imgs[idx].copy()
    score_mask_img[score_mask] = (0, 0, 255)
    score_mask_imgs[key].append(score_mask_img)

However, my attempt resulted in the images being entirely red. Could you clarify how you determined the threshold_value? Any insight would be greatly appreciated as I navigate through this learning process.

Thank you once again for your valuable contribution.