search

Newcomer520 / tf-dagmm

Tensorflow Implementation of dagmm: Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection, Zong et al, 2018
56 stars 17 forks source link

For which kind of image sizes is this suitable? #2

Open Santana18 opened 6 years ago

Santana18 commented 6 years ago

Are modifications necessary for larger camera images? Does your code expect any multiples of image sizes?

The network worked well for you, according to the paper content?

Thanks for your implementation and the good readme.

Newcomer520 commented 6 years ago

Hello, Thanks for your interesting of this repository. In my experiments I took sizes of (128, 128) and (64, 64) for the input. The configuration could be customized for the specific image size and filters of the autoencoder. I used the basic autoencoder(the auther did as well) to do the anomaly detection and it looks like good for my case. However, the encoded vectors seem to contribute less to the anomaly detection. In other words, I found the key factor is still the reconstruction error in my cases. The estimated GMM did help to predict the distribution, therefore having a robust result of anomaly detection.