Is label smoothing only useful in multi-class recognition?

[irvingzhang0512]

After reading related codes, it seems label smoothing is only used in multi-class recognition case.

I think label smoothing is also useful in single-class recognition.

[irvingzhang0512]

I guess label smooth is useful in single-class case, but need some extra work(e.g. loss/pipeline) to train models?

[innerlee]

Do you have paper references (later than year 2016) about label smoothing in single class classification?

[irvingzhang0512]

Actually, my friend gave me this advice(add label smoothing in single-class classification), it works for him in some cases.

By the way, someone told me that mix-up worked in action recognition, too... If mix-up works, i'll create a pr

[irvingzhang0512]

I don't much about label smooth in action recognition. But for image classification, this paper shows that, label smoothing helps to improve ImageNet accuracy.

[kennymckormick]

In my experiments, I found label smoothing doesn't work on TSN series models. For mixup, it seems to work on training 3D models from scratch (details in https://arxiv.org/pdf/2003.13042.pdf). You can conduct experiments to see if label smoothing works for 3D models.