fixed landmark postprocessing + resize input canvases instead of tensors to net input size, which is much more performant