search

Videmo / pymot

Objective evaluation of tracker performance with the Multiple Object Tracking (MOT) measures.
127 stars 46 forks source link

Bugfix #6

Open terezasoukupova opened 7 years ago

terezasoukupova commented 7 years ago

Hello, I think I have found a bug in your code. In pymot.py, line 238, where is the overlap computed? I think that there should be something like this before this line: rect_groundtruth = Rect(groundtruths[gt_index]) rect_hypothesis = Rect(hypotheses[hypo_index]) overlap = rect_groundtruth.overlap(rect_hypothesis)

Am I right?

DenisKuplyakov commented 7 years ago

I also think it looks like a bug. But it's better to solve it this way, without intersecting rects another time:

self.total_overlap_ += 1.0 / munkres_matrix[gt_index][hypo_index]

Still surprised, that results I've get were looking adequate.

herr-biber commented 7 years ago

You are right. Thanks. Looks like a bug. I'd really appreciate a pull request and a small test example which reproduces the bug.

terezasoukupova commented 7 years ago

I am sorry, I am new to github and I do not know for what is exactly pull request useful. But another point, isn't it weird, that MOTA can be negative? E.g. when you have no misses, no mismatches and many false positives.

herr-biber commented 7 years ago

See https://help.github.com/articles/creating-a-pull-request/ on how to create a pull request. Basically you fork this repo. Commit changes to your fork and file the pull request to allow me to merge changes made in your fork to be merged into this master.

Negative MOTA is possible, yes.

DenisKuplyakov commented 7 years ago

@terezasoukupova has created pull request for this: #8