can region proposal handle a situation where one bounding box has different labels?

[ankurgupta7]

@rbgirshick i am trying to learn a network where one bbox can be say 'horse' and 'animal' at the same time. would the solver converge(meaningfully) in such a situation?

bbox1 = {x0, y0, x1, y1} -> label1
bbox2 = {x0, y0, x1, y1} -> label2
bbox3 = {x0, y0, x1, y1} -> label3
i.e bbox = {x0, y0, x1, y1} -> label1, label2, label3 ?

[guanxiongsun]

I think maybe you can just label the box with 'horse', after training, for all labeled 'horse' boxes, you can manually add a label 'animal'...Maybe..

[ankurgupta7]

Oh. I actually meant it could be horse - animal the first image and blue - animal in the second image. So they are not exactly the same. In that case I would just have had label 'horseAnimal". I think I'll edit the issue to reflect this. Thanks anyways.