how can i train faster-rcnn-end2end if one window can have multiple labels?

[ankurgupta7]

[ankurgupta7]

How can i do that?

[rohitghosh]

Hi @ankurgupta7 ,

Can you elaborate a bit more on the pain ? Ideally, while specifying annotations in 'annotations.txt' you just mention the object class along with co-ordinates for top left and bottom-right of bounding boxes for each object. Overlapping windows won't be an issue

[ankurgupta7]

Yeah, true. but i dont have a dataset that has all the labels for each image. i have some images with label a and some with b. ideally they all the images would have had labels a&b. if i train with such a dataset, loss function would penalize 'a' for images with label 'b'. because it kind of assumes that both the labels are mutually exclusive. Am i making any sense? TL;DR my dataset is incomplete. how can i not penalize labels that 'should have been ' but 'are not' present in the trainig dataset?

[ankurgupta7]

@rohitghosh my question is essentially about converting the SoftmaxLoss in "loss_cls" layer to SigmoidCrossEntropy. fails due to mismatching size between {bottom: "cls_score", bottom: "labels"}

[ycxia]

@ankurgupta7 Hi, did you solve this problem?