Utilize uncertainty for improved person-tracking/face-matching performance in Rekognition video face search

rekognition_video_face_search_uncertainty.py

import collections


def sort_raw_rekognition_results(results):
    indices = collections.defaultdict(list)  # unique people detected in video to a `list` of their `PersonMatch` objects
    
    timestamps = collections.defaultdict(list)  # `lists` maintain order so we can keep track of people and their indices
    timestamps_indices = collections.defaultdict(list)

    for p in self.results['Persons']:
        index = p['Person'].pop('Index')  # pop off indices so we identical `dicts` will equal each other
        timestamp = p['Timestamp']

        if p in timestamps[timestamp]:  # a duplicate
            i = timestamps[timestamp].index(p)
            _index = timestamps_indices[timestamp][i]

            indices[_index].remove(p)

            assert timestamps[timestamp][i] == p
            timestamps[timestamp].pop(i)
            timestamps_indices[timestamp].pop(i)

            index = ','.join(sorted([str(index), *str(_index).split(',')]))  # comma separated indices are where the uncertainty is

        assert p not in indices[index]
        assert p not in timestamps[timestamp]
        assert index not in timestamps_indices[timestamp]
        indices[index].append(p)
        timestamps[timestamp].append(p)
        timestamps_indices[timestamp].append(index)

    return indices