Is EfficientAD's argument pad_maps in the paper?

[jpcbertoldo]

The argument

anomalib/src/anomalib/models/efficient_ad/torch_model.py

Line 234 in c97762c

pad_maps=True,

in the torch model

anomalib/src/anomalib/models/efficient_ad/torch_model.py

Line 214 in c97762c

class EfficientAdModel(nn.Module):

is used at inference time (else is not self.training)

anomalib/src/anomalib/models/efficient_ad/torch_model.py

Lines 334 to 347 in c97762c

	else:
	with torch.no_grad():
	ae_output = self.ae(batch)
	map_st = torch.mean(distance_st, dim=1, keepdim=True)
	map_stae = torch.mean(
	(ae_output - student_output[:, self.teacher_out_channels :]) ** 2, dim=1, keepdim=True
	)
	if self.pad_maps:
	map_st = F.pad(map_st, (4, 4, 4, 4))
	map_stae = F.pad(map_stae, (4, 4, 4, 4))
	map_st = F.interpolate(map_st, size=(self.input_size[0], self.input_size[1]), mode="bilinear")
	map_stae = F.interpolate(map_stae, size=(self.input_size[0], self.input_size[1]), mode="bilinear")

before the bilinear interpolation.

I could not understand where is operation is referenced in the paper.

Anyone could clarify? @nelson1425 ?

[abc-125]

In the paper in the Appendix E:

In latency-critical applications, padding in the PDN architecture of EfficientAD can be disabled. This speeds
up the forward pass of the PDN architecture by 80 μs without impairing the detection of anomalies. We time EfficientAD
without padding and therefore report the anomaly detection results for this setting in the experimental results of this paper.

So you can disable the padding in PDN like this, using parameter padding :

pad_mult = 1 if padding else 0
self.conv1 = nn.Conv2d(3, 256, kernel_size=4, stride=1, padding=3 * pad_mult)

But if you do this, you need to add padding to the output of the model before resizing it, because it would be smaller than the original image (for 256x256 it would be 56x56 instead of 64x64). You can also check this answer.

[abc-125]

I did a quick overlay of this output to an original image, this way it should be more obvious what I am talking about.
With F.pad(map, (4, 4, 4, 4)), anomaly map fits the actual image:

Without it, the anomaly map is cropped closer than the original image and the bottle on it is bigger than the bottle on the image (you can see it from the green glow around the original bottle):

[abc-125]

I also checked nelson1425's implementation (with this padding) on the VisA dataset some time ago, and the results are the same as the paper.

[jpcbertoldo]

Hmm, ok I see what you mean.

In this case, it's better to automatically add this post-network paddfing when padding=False to avoid confusion/user mistake, no?

[abc-125]

In this case, it's better to automatically add this post-network paddfing when padding=False to avoid confusion/user mistake, no?

Sure, this is a good idea. It can also be worth it to rename padding to padding_conv to make it more clear which padding it is.

[jpcbertoldo]

I added it in
#1376