Add detection metrics

[jvlmdr]

This PR adds the detection metrics from the MOT Challenge. There are some differences between this implementation and the original implementation. I have outlined them here:
https://docs.google.com/document/d/1UTR4G8va_fe-KsNZhrQJNE5w_7F9fPBVy9G4dKWrBfk/preview

[jvlmdr]

I forgot to mention, the command that I used to compute detection metrics for the MOT17Det benchmark is:

python scripts/run_mot_challenge.py \
  --METRICS Det \
  --BENCHMARK MOT17Det \
  --REQUIRE_UNIQUE_IDS False \
  --MIN_VIS 0.5 \
  --DROP_OTHER_CLASSES True \
  --DROP_EXTRA_TIMESTEPS True

There are several new flags here: MIN_VIS sets the minimum visibility for the ground-truth, REQUIRE_UNIQUE_IDS disables the assertion that track IDs are unique (must be disabled for "Det" challenges), DROP_OTHER_CLASSES removes predictions with class ID > 1 and DROP_EXTRA_TIMESTEPS disables the assertion that frame IDs <= sequence length (ignoring any predictions in extra frames).

Edit: I meant --IGNORE_TRACK_IDS False instead of --REQUIRE_UNIQUE_IDS True as highlighted below in #44 (comment).

[youonlytrackonce]

Hello Jack,

in your branch: https://github.com/jvlmdr/TrackEval/tree/detection-merge
I get the following error message:
"run_mot_challenge.py: error: unrecognized arguments: --REQUIRE_UNIQUE_IDS False"

Would omitting this flag effect the results?

Thank you.

[jvlmdr]

Hey! My mistake. I forgot that I inverted the sense of this flag from --REQUIRE_UNIQUE_IDS False to --IGNORE_TRACK_IDS True. (And it should only be necessary to do this for MOT17Det where the track IDs are unused; it's not required for MOT17.)