def test_bbox_corner2center(self):
a1 = torch.FloatTensor([0, 0, 100, 100])
b1 = torch.FloatTensor([50, 50, 100, 100])
a2 = torch.FloatTensor([[0, 0, 100, 100]])
b2 = torch.FloatTensor([[50, 50, 100, 100]])
a3 = torch.FloatTensor([[50, 50, 100, 100]])
b3 = torch.FloatTensor([[75, 75, 50, 50]])
self.assertTrue(bbox_corner2center(a1).equal(b1))
self.assertTrue(bbox_corner2center(a2).equal(b2))
self.assertTrue(bbox_corner2center(a3).equal(b3))
def _refine_proposal(self, proposal, bbox_reg):
"""Refine proposal bbox with the result of bbox regression.
Args:
proposal(Tensor): (x1, y1, x2, y2), bbox proposal from RPN.
bbox_reg(Tensor): (dx, dy, dw, dh), bbox regression value.
Returns:
bbox_refined(Tensor): (x1, y1, x2, y2)
"""
x, y, w, h = bbox_corner2center(proposal).chunk(4)
dx, dy, dw, dh = bbox_reg.chunk(4)
px, py = w * dx + x, h * dy + y
pw, ph = w * torch.exp(dw), h * torch.exp(dh)
bbox_refined = bbox_center2corner(torch.cat([px, py, pw, ph]))
px1, py1, px2, py2 = bbox_refined.chunk(4)
px1 = torch.clamp(px1, max=self.img_width, min=0)
px2 = torch.clamp(px2, max=self.img_width, min=0)
py1 = torch.clamp(py1, max=self.img_height, min=0)
py2 = torch.clamp(py2, max=self.img_height, min=0)
bbox_refined = torch.cat([px1, py1, px2, py2])
return bbox_refined
def _get_bbox_targets(proposals, gt_bboxes):
""" Calculate bounding box targets, input coord format is (left, top, right, bottom),
see R-CNN paper for the formula.
Args:
proposals(Tensor): [n, 4]
gt_bboxes(Tensor): [n, 4]
Returns:
bbox_targets(Tensor): [n, 4]
"""
proposals = bbox_corner2center(proposals)
gt_bboxes = bbox_corner2center(gt_bboxes)
xy = (gt_bboxes[:, :2] - proposals[:, :2]) / proposals[:, 2:]
wh = torch.log(gt_bboxes[:, 2:] / proposals[:, 2:])
x, y = xy.chunk(2, dim=1)
w, h = wh.chunk(2, dim=1)
bbox_targets = torch.cat([x, y, w, h], dim=1)
return bbox_targets