About the meaning of "n" of "DDAD-S-n"

[xgzng]

Hello, I'd like to ask about the meaning of 10 in DDAD-S-10 in your paper. Your paper describes it as "n refers to the number of denoising iterations", but isn't the number of denoising iterations 1000? I don't see a 10 setting in the code, or do I iterate the training process 10 times to get it?

[arimousa]

Hello, as you correctly mentioned 10 is the number of the denoising steps. In the code it can be calculated based on "test_trajectoy_steps" and "skip" in the config file. Where "test_trajectoy_steps: 250" means the starting point (T' = 250) and "skip" means the denoising step size we take at each denoising iteration. Mathematically n= test_trajectoy_steps / skip.

[xgzng]

Thank you for your answer.I'll take a closer look at the code.

[arimousa]

We utilize the DDIM sampler which allows for the generation of samples without the need to take all 1000 steps. In our model, during inference, the initial noisy step is 250 (referred as test_trajectoy_steps in the config file) which is a more suitable noise-to-signal ratio.

Note that during training, all 1000 steps (referred as trajectory_steps) are executed.

[xgzng]

Hello, I visualized the reconstructed image in the reconstruction part of your code, but I found that the reconstructed image did not have much repair, and there was still a big anomaly. I did the above operation in the bottle dataset. Could you give me a solution?

[arimousa]

Decreasing the Value of w would result in better reconstruction of the anomalous part.

[xgzng]

OK, thank you for your answer, it does have a very good effect, but it does not seem to achieve the most perfect repair effect. It still feels flawed to the naked eye, but it is enough for anomaly detection, perhaps because the abnormal part is too serious to achieve an almost perfect repair effect. I only made observations on the bottle dataset.

[arimousa]

Great. If your are using DDAD-S just scaling the model to higher number of parameters will help. However, though we aimed for a cohesive visual reconstruction, it is worth to note that at the end the anomaly score is calculated based on the deviation of the reconstruction from the input image. It means that only keeping then normal pattern of the input image, while having deviation in anomalous part will result in anomaly detection. The conditioning mechanism essentially aims for keeping the normal behavior of the image.
Hope this explanation helps to map out the correlation between reconstruction and anomaly detection.

[xgzng]

Hello, I would like to ask, will DDAD have a significant improvement over DDAD-S in the effect of reconstruction?

[arimousa]

Hi, DDAD can reconstruct details with greater precision, resulting in a better anomaly score.

[xgzng]

OK, thank you very much.

[xgzng]

Hello, is there a range limit for the conditional parameter w? For example, when w takes a certain maximum value, the effect of the diffusion model will be almost zero