IMPORTANT: Be sure to replace all template sections {{ like this }} or your issue may be discarded.

Overview

{{ Hi, I have noticed that there are not usage example of how to get the motion vector, and the only clue I have got lies here in the unittest. Could you please show a standard and efficient usage of transfer a motion vector into a mv map? Namely the motion offset of each pixel, a matrixs of size (2, height, weight)?

A simple implementation can be this, but it use a nested loop on each pixel, seems inefficient. }}

Desired Behavior

{{

    def test_decoded_motion_vectors(self):

        container = av.open(fate_suite('h264/interlaced_crop.mp4'))
        stream = container.streams.video[0]
        codec_context = stream.codec_context
        codec_context.options = {"flags2": "+export_mvs"}

        for packet in container.demux(stream):
            for frame in packet.decode():
                vectors = frame.side_data.get('MOTION_VECTORS')
                vectors_map = frame.side_data.parse_mv(vectors)# TODO
                if frame.key_frame:
                    # Key frame don't have motion vectors
                    assert vectors is None
                else:
                    assert len(vectors) > 0
                    return

}}

Example API

{{
def parse_mv(vector) -> np.ndarray:
Args:
    vector: np.ndarray, [(-1, src_x, src_y, dst_x, dst_y, motion_x, ...), ()..]
Returns:
    np.ndarray: of shape (2, h, w) 
}}

Additional context

{{ Add any other context about the problem here. }}