This updates the AnchorTarget() layer to function with inputs of size (None, None, num_channels).
The unmap() function has been completely removed. Instead, anchors outside the image are retained when calculating overlaps, then their labels are overwritten afterwards.
The signature of the label and overlapping functions are now consequently simpler. For example, overlapping(anchors, gt_boxes, inds_inside) is now simply def overlapping(anchors, gt_boxes), which is a lot more intuitive.
Calls to keras.backend.int_shape are gone, so no more errors when input sizes are None.
yhenon
commented
6 years ago
This updates the
AnchorTarget()layer to function with inputs of size(None, None, num_channels).unmap()function has been completely removed. Instead, anchors outside the image are retained when calculating overlaps, then their labels are overwritten afterwards.labelandoverlappingfunctions are now consequently simpler. For example,overlapping(anchors, gt_boxes, inds_inside)is now simplydef overlapping(anchors, gt_boxes), which is a lot more intuitive.keras.backend.int_shapeare gone, so no more errors when input sizes are None.