code for processing a single picture

[kanwangshijie]

Hello, thank you very much for opening up the source code. When reading your paper recently, I found that your method can detect whether a single picture is abnormal. But when I read your source code, I found that they are all used for video detection. I now want to use only the reconstruction part to detect whether a single image is abnormal. There is no need to extract box, flow and other preprocessing operations, just like you deal with the Minist data set in your paper, so I would like to ask you to provide the code for detecting the Minist data set?

您好，十分感谢您开源了源代码，最近在拜读您的论文时，发现您的方法可以检测单张图片是否异常。但是我阅读您的源代码时，发现都是用于用于视频检测的。我现在想只使用重建部分检测单张图片是否异常，不需要提取box、flow等预处理操作，就像您在论文里处理Minist数据集那样，因此我想问您提供了检测Minist数据集的代码了吗？

[LiUzHiAn]

The images are pretty much similar to the videos, you can try to adapt the Memory-augmented autoencoder part to your own task.

[bing-0906]

The images are pretty much similar to the videos, you can try to adapt the Memory-augmented autoencoder part to your own task.

Hi. Thanks for your excellent work. I would like to reimplement this work for online video stream. I have some questions during learning the code and the paper. In your paper, it shows that it will take 100ms for one frame. Does it include the time for model loading? Is it possible for improving the efficiency further? As we know, online processing requires 25fps at least.

[bing-0906]

The images are pretty much similar to the videos, you can try to adapt the Memory-augmented autoencoder part to your own task.

Hi. Thanks for your excellent work. I would like to reimplement this work for online video stream. I have some questions during learning the code and the paper. In your paper, it shows that it will take 100ms for one frame. Does it include the time for model loading? Is it possible for improving the efficiency further? As we know, online processing requires 25fps at least.

I tried some other VAD framework, such as method based on skeleton detection. The time cost for model loading and object detection is huge, which is also a step for your method. Maybe this is not your research purpose. I just wonder if you have better idea. ^^ look forward to your response

[LiUzHiAn]

Hi, the model loading time was not included. Since we follow the preprocessing steps in this paper, you can try some modifications to improve efficiency, such as using a lightweight object detector, considering some techniques other than gradient to extract some complementary foreground objects. But this may have a certain impact on the performance.