Right now, the CNN's negatives training set only includes patches with no overlap with the GT at all. That's a mistake; I should be including more challenging negatives which include the GT, but with a low overlap so that the CNN learns not to activate on off-centre detections (kill some of its spatial invariance!). These aren't "hard negatives" in the traditional sense of "misclassifications produced by a previous version of the detector", but rather just challenging negatives. I think that real hard negatives would be too difficult to implement in the time left.
Right now, the CNN's negatives training set only includes patches with no overlap with the GT at all. That's a mistake; I should be including more challenging negatives which include the GT, but with a low overlap so that the CNN learns not to activate on off-centre detections (kill some of its spatial invariance!). These aren't "hard negatives" in the traditional sense of "misclassifications produced by a previous version of the detector", but rather just challenging negatives. I think that real hard negatives would be too difficult to implement in the time left.