def forward(self, features, rois):
#need to normalised (x0,y0,x1,y1) to [0,1]
height, width = features.size()[2:4]
ids, x0, y0, x1, y1 = torch.split(rois, 1, dim=1)
# ids, x0, y0, x1, y1,_,_= torch.split(rois, 1, dim=1)
ids = ids.int()
x0 = x0 * self.spatial_scale
y0 = y0 * self.spatial_scale
x1 = x1 * self.spatial_scale
y1 = y1 * self.spatial_scale
if 0:
#https://github.com/ppwwyyxx/tensorpack/issues/542
#https://github.com/ppwwyyxx/tensorpack/blob/6d5ba6a970710eaaa14b89d24aace179eb8ee1af/examples/FasterRCNN/model.py#L316
#-0.5 because bilinear sample assumes floating point coordinate (0.0, 0.0) is the same as pixel value (0, 0))
scale_x = (x1 - x0) / (self.crop_width)
scale_y = (y1 - y0) / (self.crop_height)
nx0 = (x0 - 0.5 + scale_x / 2)
ny0 = (y0 - 0.5 + scale_y / 2)
nw = scale_x * self.crop_width
nh = scale_y * self.crop_height
nx1 = nx0 + nw
ny1 = ny0 + nh
nx0 = nx0 / (width - 1)
ny0 = ny0 / (height - 1)
nx1 = nx1 / (width - 1)
ny1 = ny1 / (height - 1)
boxes = torch.cat((ny0, nx0, ny1, nx1), 1)
if 1:
x0 = x0 / (width - 1)
y0 = y0 / (height - 1)
x1 = x1 / (width - 1)
y1 = y1 / (height - 1)
boxes = torch.cat((y0, x0, y1, x1), 1)
boxes = boxes.detach().contiguous()
ids = ids.detach()
return CropAndResizeFunction(self.crop_height,
self.crop_width,
extrapolation_value=0)(features, boxes,
ids)
def forward(self, features, rois):
#need to normalised (x0,y0,x1,y1) to [0,1]
height, width = features.size()[2:4]
ids, x0, y0, x1, y1 = torch.split(rois, 1, dim=1)
ids = ids.int()
x0 = x0 * self.spatial_scale
y0 = y0 * self.spatial_scale
x1 = x1 * self.spatial_scale
y1 = y1 * self.spatial_scale
if 1:
x0 = x0 / (width - 1)
y0 = y0 / (height - 1)
x1 = x1 / (width - 1)
y1 = y1 / (height - 1)
boxes = torch.cat((y0, x0, y1, x1), 1)
boxes = boxes.detach().contiguous()
ids = ids.detach()
return CropAndResizeFunction(self.crop_height,
self.crop_width,
extrapolation_value=0)(features, boxes,
ids)
def forward(self, image, boxes, box_ind):
return CropAndResizeFunction(self.crop_height, self.crop_width, self.extrapolation_value)(image, boxes, box_ind)