Pred_mask coming out blank, even though the heat-map and anomaly-maps show the anomaly being detected.

[LP-Isaac]

Hey guys, I'm with MacnicaDHW and I've been having some issues while inferencing on a custom dataset with only normal_dir.

The issue: The pred_mask comes out empty, with seemingly no anomalies shown. The scores are all pretty great, and the heat and anomaly maps are very promising. The issue is that the pred_mask comes out completely blank.

I've tried adding huge anomalies to my test image as a simple test to check whether it was a threshold issue, and a few times this resulted in am anomaly showing in the pred_mask, though it was much smaller than it should've been.

From this, I thought that it could really be a threshold issue, and tried inferencing with different image and pixel thresholds with this addition:

model.adaptive_threshold = False
model.pixel_threshold.value = torch.tensor(float(1))
model.image_threshold.value = torch.tensor(float(1))

Where I tested a wide range of options to obersve the different outcomes. This resulted in changes as far as the anomaly and heat maps are concerned, but no progress on the pred_masks.

I've also noticed that my pred_score seems quite low (0.3775), and the pred_label comes out as False. I whatever is making the result False is the cause of the issue?

Dome details:

Model: FastFlow

Dataset: Folder, using only normal_dir and TestSplitMode.SYNTHETIC

Python: 3.10.14

Anomalib: 1.1.0dev

Please let me know if there's anything else I should provide to make your jobs easier.

Thanks in advance for the time and help!

[TurboJonte]

I've had something similar happened, for some images when they've been labeled as "ABNORMAL" there was just an empty mask. Does it happen for every image in your case?

[blaz-r]

I think that this indeed is an issue with thresholding. Note that image scores undergo a separate thresholding and normalization process from anomaly maps, so it can happen that there is a discrepancy.
One additional thing, if you use normalization, the scores as well as masks are normalized to a range around [0, 1], but Anomalib also takes the threshold into account for this normalization.

You could try to tune the threshold, but I'd recommend that you first check the values produced by automatic threshold.