def forward(self, fmap, rois, union_inds):
union_pools = union_boxes(fmap, rois, union_inds, pooling_size=self.pooling_size, stride=self.stride)
if not self.use_feats:
return union_pools.detach()
pair_rois = torch.cat((rois[:, 1:][union_inds[:, 0]], rois[:, 1:][union_inds[:, 1]]),1).data.cpu().numpy()
# rects_np = get_rect_features(pair_rois, self.pooling_size*2-1) - 0.5
x1_union = np.minimum(pair_rois[:, 0], pair_rois[:, 4])
y1_union = np.minimum(pair_rois[:, 1], pair_rois[:, 5])
x2_union = np.maximum(pair_rois[:, 2], pair_rois[:, 6])
y2_union = np.maximum(pair_rois[:, 3], pair_rois[:, 7])
w = x2_union - x1_union
h = y2_union - y1_union
w0 = np.where(w==0)[0]
h0 = np.where(h==0)[0]
if w0.shape[0]>0:
print ('w==0 at ', w0)
if h0.shape[0]>0:
print ('h==0 at ', h0)
pair_rois[np.where(w==0), 6] = x1_union[np.where(w==0)] + 1.
pair_rois[np.where(h==0), 7] = y1_union[np.where(h==0)] + 1.
rects_np = draw_union_boxes(pair_rois, self.pooling_size*4-1) - 0.5
rects = Variable(torch.FloatTensor(rects_np).cuda(fmap.get_device()), volatile=fmap.volatile)
if self.concat:
return torch.cat((union_pools, self.conv(rects)), 1)
return union_pools + self.conv(rects)
def forward(self, fmap, rois, union_inds):
union_pools = union_boxes(fmap,
rois,
union_inds,
pooling_size=self.pooling_size,
stride=self.stride)
if not self.use_feats:
return union_pools.detach()
pair_rois = torch.cat(
(rois[:, 1:][union_inds[:, 0]], rois[:, 1:][union_inds[:, 1]]),
1).data.cpu().numpy()
# rects_np = get_rect_features(pair_rois, self.pooling_size*2-1) - 0.5
rects_np = draw_union_boxes(pair_rois, self.pooling_size * 4 - 1) - 0.5
rects = Variable(torch.FloatTensor(rects_np).cuda(fmap.get_device()),
volatile=fmap.volatile)
if self.concat:
return torch.cat((union_pools, self.conv(rects)), 1)
if conf.rl_offdropout:
self.conv[2].training = False
self.conv[6].training = False
return union_pools + self.conv(rects)
def forward(self, fmap, rois, union_inds, use_feats=True):
union_S_pools, union_O_pools = new_union_boxes(
fmap,
rois,
union_inds,
pooling_size=self.pooling_size,
stride=self.stride)
if use_feats:
pair_rois = torch.cat(
(rois[union_inds[:, 0]], rois[union_inds[:, 1]]),
1).data.cpu().numpy()
# rects_np = get_rect_features(pair_rois, self.pooling_size*2-1) - 0.5
rects_np = draw_union_boxes(pair_rois,
self.pooling_size * 4 - 1) - 0.5
rects = Variable(torch.FloatTensor(rects_np).cuda(
fmap.get_device()),
volatile=fmap.volatile)
return union_S_pools, union_O_pools, self.conv(rects)
else:
return union_S_pools, union_O_pools, None