About Loss about hed-unet HOT 5 CLOSED

from hed-unet.

That is odd, my assumption would be that the auto-weighting loss is a bit out of control. Here's a possible fix. Try replacing /deep_learning/loss_functions with this adapted file, which includes a hybrid loss: https://gist.github.com/khdlr/c0bb6a653ac00c6d6859e6dc773daf96

Then select the hybrid loss function by changing lines 10-11 in config.yml to read

loss_args:
  type: HybridLoss

Of course there's no way of knowing whether it will actually help, but it might just do the trick :)

from hed-unet.

Thanks.
I want to design or use a loss function similar to box loss and giou loss in target detection. Because the object edge detection of my task is always closed. At present, some gaps and edges will lead to false edge detection. After using the above losses, it may be alleviated.

from hed-unet.

I want to know whether this job can train negative samples close to positive samples (labels may be set to be all empty). For example, I train unet-hed to identify the edge of white paper, but some white rectangular objects will also be mistakenly detected as white paper. Is this a common problem of segmenting networks? Or can it be fitted by negative sample training? What to do.

from hed-unet.

Yeah, this should be easy to fix by adding negative samples to the training set like you suggested.

from hed-unet.