About K value when calculating loss

[VikingKang]

Thanks for your sharing.
I found that when you calculate the loss, the K you use seems to be based on batchsize, not a single image. Did you decide to do this based on relevant considerations? For example, there is little difference between batchsize and single image.
Maybe it is better if I calculate K separately for each picture?

for your paper:
L = Ldetection + αLclassification (1)
Detection loss Ldetection itself is a weighted sum of three components: mean binary crossentropy loss on positive pixels Lp, mean binary crossentropy loss on negative pixels Ln, and mean binary crossentropy loss on worst predicted k negative pixels Lh, where k is equal to the number of positive pixels in image.

[asmekal]

If we choose K among batch the average loss for this loss term will be higher than if we choose K for each image separately, so we will have stronger penalty on hard negatives. You can imagine that most of the images in batch have only 'easy negatives' which will not impact the loss significantly and therefore it might be better to pay more attention to other images in a batch. This was our motivation, so we haven't tried separate image negatives selection

Thanks for sharing the citation from the paper. Yes, we used batch-K selection, not a separate selection for each image, the description in the paper is incorrect here :(