After further inspection, I’ve noticed that the only function of the IDMAP in kittiTrack2mot.sh is to include or discard classes when building the <video_name>.txt files at each checkpoint with bounding boxes, and annotations.
No class ID is added in any way by the script, as it just builds those MOT annotation files with:
printf "%u,%d,%d,%d,%d,%d,%f,-1,-1,-1\n" \
"$frame" "$id" "$x1" "$y1" "$width" "$height" "$confidence" >> $OUT_FILE
If the output files are the only ones used in the evaluation step, I guess that having the class ID included in the dataset as mentioned in the documentation is actually doing nothing. I know that metrics like MOTA don’t take classes into account, but I was expecting for PipeTuner to use them for the tuning of the PGIE.
By just setting to 1 the class ID fields in all the dataset, plus with the proposed modification in kittiTrack2mot.sh, everything is working fine (unless there are extra features that would take advantage of knowing about classes).
So if this is expected behaviour, the problem is fixed, but the documentation is a little misleading regarding the dataset building.