Does the anchor should be different since mosiac aug make whole box distribution changed?

[lucasjinreal]

Does the anchor should be different since mosiac aug make whole box distribution changed?

[lucasjinreal]

Currenly, I only saw the anchor box is absolute and fixed rather than relative to original w,h.

[lucasjinreal]

@glenn-jocher thank u for your reply. I understand autoanchor generates anchors automatically. But I want try another strategy:
Say I want train it with original image size without specifiying a fixed --img-size, if a batch images with different sizes, I resized it to the largest one. So in this siutation, original anchor based on original dataset's img sizes still works?

Since the img resized, so actually the targets larger in terms of absolute w and h. what do u think?

[glenn-jocher]

@jinfagang --rect will dynamically size batches on the short dimension as you describe. In all cases --img will define the long side of a batch. You would have to customize the dataloader to implement your idea, this is not available off the shelf currently.

[lucasjinreal]

@glenn-jocher thanks for your suggestion. Do u think if keep anchor ratio but dynamic make it grow with input shape will work from theoretically？

[glenn-jocher]

@jinfagang you have to experiment to see what works best for you custom data and model combination.

[lucasjinreal]

@glenn-jocher In my opinion, I think it won't make model converge. since anchor changes every single batch, when inference, it hardly predict exactly location of objects.