how to deploy this method in an industry application

[mrljwlm]

can this model export to tensorRT or onnx format? if I want to use it in an industry application, any idea?

[DeepKnowledge1]

You can use it directly

[bjaeger1]

This PaDiM model cannot be exported to TensorRT or ONNX, as there is no "real" model training, hence no trained CNN model which has to be exported or optimized. The CNN in this approach is only(!) used for feature maps extraction. The PaDiM approach can be implemented to be as fast (or faster in my case) than other anomaly detection approaches which use CNN (autoencoders, etc) for model training and inference.

[RichardChangCA]

In practice, how to define the decision threshold to get precision and recall? In this paper, the author find the optimal threshold to achieve the best f1 score based on the precision recall curve from all testing data. In the real world, we do not have the sufficient test data to get the decision threshold, how can we deal with this issue? Thanks