Xincheng Yao
Xincheng Yao
As the log-likelihhoods theoretically may be in (-∞，0)，in practice, we found the log-likelihoods are some large negative values. So, we use a normalizer to reduce the values to about -1...
Can you give me the detailed error line? This error may occurr in the dataset codes.
You can try to use the plot_visualizing_results function in the utils/visualizer.py
I think you can consider that the abnormal and normal boundaries are used as contrastive object in our BG-SPP loss. After replacing the logo term with pair-wise distances, our BG-SPP...
You can organize your dataset as the MVTecAD dataset format. And change the root directory to your dataset. Or you can see the dataset class in the mvtec.py, and try...
Sorry. I can't find the previous checkpoints. You can try to run the code from scratch.
The foreground mask images are in fg_mask.zip file, you can unzip this file.
Sorry, I didn't try how to convert the model to onnx. But converting pytorch model to onnx should be feasible.
I have no idea what data leakage means. We use few-shot abnormal samples (~2) for each anomaly type and increase the diversity of generated abnormal samples through random augmentation.
1. 是的，fg_mask用于提供前景区域，避免缺陷生成在背景区域。 2. 是先生成边界再进行边界引导，但在代码中是在每个batch中动态确定边界。 3. 异常样本应该是越多越好，看你总共有多少个异常样本可用，我们在mvtec上选择10个，是基于异常稀缺性的假设。