init_label_function for classification online train about pytracking HOT 3 CLOSED

Is the code "self.pos - self.pos.round()" means to transfer the original point from left-up to self.pos.round() ?

from pytracking.

In code "center_pos = sz * target_center_norm + 0.5 * torch.Tensor([(ksz[0] + 1) % 2, (ksz[1] + 1) % 2])"

"sz * target_center_norm" part transfer the normalized center pos to feature map coordinate, what about another part " + 0.5 * torch.Tensor([(ksz[0] + 1) % 2, (ksz[1] + 1) % 2])" 's mean?

and I find after modify from ”center_pos = sz * target_center_norm + 0.5 * torch.Tensor([(ksz[0] + 1) % 2, (ksz[1] + 1) % 2])" to "center_pos = sz * target_center_norm", when do tracking, it also work well.

from pytracking.

Hi.
Just pushed an update that should clarify things a little. There was a small bug that could lead to 0.5 pixel misalignment.

self.pos.round() is the center pixel coordinate of the extracted sample. So self.pos - self.pos.round() should correct from this sub-pixel displacement.

The term 0.5 * torch.Tensor([(ksz[0] + 1) % 2 is to correctly handle even and odd kernels. It leads to small misalignments if you remove that, but the IoU-Net largely compensates for it.

from pytracking.