Hiya,

1. Yes, I think that the bigger the blob the easier for it to be separated, as networks often struggle with small resolutions.
2. The scale in the global split loss is indeed the number of points.
3. You are right that there is imbalance between positive and negative samples. It would be interesting to optimize the focal loss to handle that.

from lcfcn.

I read the paper again. the blob will be remove just retain the blob contains point .
the probs will drop down and can not form a blob. pretty good!

from lcfcn.

split_mode loss

if blob_dict["n_multi"] > 0:
    loss += compute_split_loss(S_log, S, points, blob_dict)

# Global loss 
S_npy = ut.t2n(S.squeeze())
points_npy = ut.t2n(points).squeeze()
for l in range(1, S.shape[1]):
    points_class = (points_npy==l).astype(int)

    if points_class.sum() == 0:
        continue

    T = watersplit(S_npy[l], points_class)
    # imsave(batch["name"][0], T*255)
    T = 1 - T
    # print(batch["image_path"][0])
    scale = float(counts.sum())
    loss += float(scale) * F.nll_loss(S_log, torch.LongTensor(T).cuda()[None],
                    ignore_index=1, reduction='elementwise_mean')

what is the difference between the split loss and globa loss?
how the scale works?
it will narrow the blob?
please tell me!

from lcfcn.

Hi there!

1. what is the difference between the split loss and global loss?

The global loss is just another variation of the split loss. There is a local split loss and a global split loss.

• The local split loss makes the model set the split boundaries that only separate between objects that are in the same blob as background.
• The global split loss makes the model set the split boundaries that separate all objects in the image as background.

2. how the scale works?

The model automatically learns the scale. It will find the right scale that allows it to output one blob per object.

3. it will narrow the blob?

Yes, without the split loss the model will output one big blob containing many objects. With the split loss, the model will output smaller blobs to make sure there is one blob per object.

from lcfcn.

yeah I figure out the the split loss.
the scale in local split loss is the num of pixs in blob. can I deem the bigger the blob is .the easier the blob will be separate? and the scale in global split loss is the num of the points.
my thoughts may not be mature enough. whether there is a problem between the ratio between positive samples and negative samples? the num of negative samples is bigger than positive samples.is helpful if use focal loss?
thanks for your detailed and fast explaination.

from lcfcn.