Inconsistent bbox annotations in COCO format annotation file and gt.txt about motsynth-baselines HOT 2 CLOSED

Hi,

Sorry for my late reply. area refers to the segmentation mask area of the given object, you can compute it yourself from the provided RLE mask using pycocotools.

We still include boxes with area= 0(i.e. totally invisible) within trajectories in both the .json COCO format annotation files and MOT format gt.txt files. This is done for consistency with other MOTChallenge datasets (MOT15, MOT16, MOT17, and MOT20). Note that boxes with area=0 will have their visibility set to 0 (9th column in MOT format files)!

The reasoning is that even though a box might be invisible for a few frames due to occlusion, a good multi-object tracker should be able to recover its rough location by applying linear interpolation or a motion model to the visible segments of the trajectory.

Best,

Guillem

from motsynth-baselines.

Hi,

Sorry for my late reply. area refers to the segmentation mask area of the given object, you can compute it yourself from the provided RLE mask using pycocotools.

We still include boxes with area= 0(i.e. totally invisible) within trajectories in both the .json COCO format annotation files and MOT format gt.txt files. This is done for consistency with other MOTChallenge datasets (MOT15, MOT16, MOT17, and MOT20). Note that boxes with area=0 will have their visibility set to 0 (9th column in MOT format files)!

The reasoning is that even though a box might be invisible for a few frames due to occlusion, a good multi-object tracker should be able to recover its rough location by applying linear interpolation or a motion model to the visible segments of the trajectory.

Best,

Guillem

Understood. Thanks for your clarification!

from motsynth-baselines.