Negative bounding box values in results

[mkoshkina]

I am testing with MOT17:

python test.py tracking --exp_id mot17_half --dataset mot --dataset_version 17halfval --track_thresh 0.4 --resume --is_recurrent --gru_filter_size 7 --num_gru_layers 1 --visibility_thresh_eval 0.1 --stream_test --only_ped --ltrb_amodal --visibility --gpus -1 --save_results --load_model ../models/mot_half.pth

I am seeing negative bounding box values in the results: 1,1,1071.49,-255.11,301.49,918.18,-1,-1,-1,-1 1,2,-12.69,655.29,114.61,344.20,-1,-1,-1,-1 1,3,1773.94,516.73,34.75,2.88,-1,-1,-1,-1

Is this expected? If so, how do you suggest we deal with those? Thank you.

[pvtokmakov]

Hi,

MOT is annotated with amodal bounding boxes, that is, if a part of a person falls outside of the frame the model still needs to predict a box for the whole body. If you only need a box for the visible fraction of the person, you can trim the values to the dimensions size (set to zero for negative values).