search

Jumabek / convert_MOT16_to_yolo

Converts MOT16 and MOT17 annotations to YOLO format.
17 stars 4 forks source link

negative number and empty file of MOT-02 #1

Open rickythink opened 7 years ago

rickythink commented 7 years ago

Much thanks for sharing the yolo format label.

I have got two questions:

  1. some of the nums in MOT17-04 contains negative number, is that right? for example: https://github.com/Jumabek/convert_MOT16_to_yolo/blob/9add042d83edf872691418274f0131202c55a9e6/sample_annotations_you_should_get/MOT17-04/labels/000004.txt#L24

  2. the whole MOT17-02's files are empty, seems that you haven't uploaded the final result, are u?.

Jumabek commented 6 years ago

Hi @RickyWong33 , sorry I just saw this.

  1. I will verify this until next week.
    1. Yes you right I forgot to generate this.
spinoza1791 commented 6 years ago

How did you customize YOLO to ignore the (-1) labels?

Jumabek commented 6 years ago

I do not remember exactly, However, here is my suggestion.

inside region_layer.c when computing the delta's for each detection bounding box and their class scores. check if the ground truth box (which is person) that has maximum IoU with predicted box has a label=-1. If so, simply do not compute the deltas for class Scores and regions. (That delta, is later used for computing the cost/loss)