Wrong results

[PasaImage]

Wrong results

[jvlmdr]

This is probably caused by the non-pedestrian objects that are present in the MOT17 ground-truth, which have confidence set to 0. The apps.eval_motchallenge script filters the ground-truth by min_confidence while the same filtering is not applied to the predicted tracks. See the difference here:

py-motmetrics/motmetrics/apps/eval_motchallenge.py

Lines 101 to 102 in 008f45c

	gt = OrderedDict([(Path(f).parts[-3], mm.io.loadtxt(f, fmt=args.fmt, min_confidence=1)) for f in gtfiles])
	ts = OrderedDict([(os.path.splitext(Path(f).parts[-1])[0], mm.io.loadtxt(f, fmt=args.fmt)) for f in tsfiles])

As a result, the non-pedestrian objects will be present in the predictions but not the ground-truth (100% recall but < 100% precision, as you are seeing).

This implementation matches the MOT challenge documentation: https://motchallenge.net/instructions/

The conf value contains the detection confidence in the det.txt files. For the ground truth, it acts as a flag whether the entry is to be considered. A value of 0 means that this particular instance is ignored in the evaluation, while any other value can be used to mark it as active. For submitted results, all lines in the .txt file are considered.