❓ EfficientAD does not support input sizes smaller than 256.

[yjszyd0813]

What is the motivation for this task?

I have checked the structure of EfficientAD and it does not support input sizes smaller than 256. Would you like to consider improving this issue? This is caused by the Encoder in AutoEncoder. In practical applications, there may be images with dimensions smaller than 256.

Describe the solution you'd like

Additional context

No response

[blaz-r]

Hello. I believe that this is sadly not viable. The structure is defined by the authors of EfficientAD, so changing it would probably affect the performance.

[yjszyd0813]

Thank you for your reply. I understand what you mean. I think if we maintain the same approach as the author and redesign a network structure, we should also achieve good performance. Do you think this can be successful?

[blaz-r]

I think it could work, probably maintaining some of the good performance, but I can't say for sure.

[watertianyi]

Is there any solution to this now?

[Likhith-Sugganahalli]

@watertianyi @blaz-r

In cases like this, would centre padding the image with (0,0,0) values not work?

[alexriedel1]

@watertianyi @blaz-r

In cases like this, would centre padding the image with (0,0,0) values not work?

sure, but the easiest way would be to resize your images to 256,256
there is also a good chance that this will improve your prediction accuracy