search

zhangzjn / GPT-4V-AD

Code for "Exploring Grounding Potential of VQA-oriented GPT-4V for Zero-shot Anomaly Detection"
27 stars 1 forks source link

How to calculate "score" with GPT-4V #2

Open z-stefanie opened 7 months ago

z-stefanie commented 7 months ago

Hello! I am a university student in Beijing, China, and I am also doing research on multi-modal large model anomaly detection. I am very interested in your work! You gave the "rigion-division" method here, I managed to reproduce it and it's great!

However, I don't know how to calculate "metrics" with GPT-4V. I refer to the No.15 reference(WinClip) of your article. He used "WinClip model" to calculate "score"(score = model(data)). Then he successfully obtain "metrcis"(result_dict = metric_cal(np.array(scores)), gt_list, gt_mask_list, cal_pro=cal_pro)).

But I truly don't know how to use "GPT-4V+Prompt+rigion-division" or other methods to calculate the "score"(I know that as long as I can calculate the "score", "metrics" will come naturally). Can you give me some guidance?

This is my first time on github to submit issue, please forgive me if I'm offended. Thank you very much indeed!

zhangzjn commented 7 months ago

The mask is generated as the anomaly score that is binary.

call-me-akeiang commented 1 week ago

Hi, boy, do you solve the problem? @z-stefanie