1.why featmap_stride = 0.4? should it be 8? cooresponding to downsampling 3 times。
in function gen_sample_grid:
you calculate the center points of (4,7) region
xx = torch.linspace(-.5, .5,window_size[0]).type_as(box).view(1, -1) wg
yy = torch.linspace(-.5, .5,window_size[1]).type_as(box).view(1, -1) lg
xx and yy not the distance between part center and object's center
for example : window is (2,2) xx = (-0.5wg, 0.5wg)
the part center x is (-.0.25wg, 0.25wg)
about the final loss:
I see in function def parse_losses(losses):
you add all loss, losses is a dict, for each loss in the dict, you should already multiply the weight for the loss, you set u= 0.9, lambda = 2 in paper, but it looks like in the code , the weights not occur:
in def aux_loss(self, points, point_cls, point_reg, gt_bboxes):
u and lambda not occurs
1.why featmap_stride = 0.4? should it be 8? cooresponding to downsampling 3 times。