USV-based Embedded Obstacle Segmentation - No segment areas

[Mateus Raitz]

I see that in the LArs dataset there is no a background class. But the boat or USV where the camera is mounted does not have a class. So when training my model, it will always segment that part which is not supposed to segment. I attached images of an example. The left image is the label and the right is the model's prediction. Is that normal or is there something in the training configs I got wrong?  Is that related to reduce_zero_labels or ignore_idx? We used the mmsegmentation-macvi as a reference but with a newer version of mmsegmentation, so our config is a little bit different than mmsegmentation-macvi.

Best regards,

Mateus Raitz.

[Lojze Žust]

Hi Mateus,

Some parts of the image in LaRS are labeled as Ignore. Anything predicted inside these regions will not be evaluated. However often, these regions are still predicted as obstacles.

Judging from your images, maybe you trained your method to predict "water" on these regions instead of ignoring them, which might hurt the performance of the method a little bit.

ignore_index tells the method which class ID to ignore during training. For LaRS that is 255 (labels are explained here).

reduce_zero_labels is used when background class uses label ID 0. In LaRS this is not the case so this should be False.

I hope this answers your questions.

Best,

Lojze Zust

petek, 20. december 2024 ob 19:09:32 UTC+1 je oseba Mateus Raitz napisala: