trajectory_dataset.py issues when training on MOT17

[simon-rob]

I have noticed a couple of issues:

1. The Dataloader ( _load_data, _load_image_anns) is internally loading images (img = cv2.imread(img_path)) which are never used in the __getitem__ function therefore dramatically increasing the overhead of the Dataloader.

2. No bounding boxes are being returned by __getitem__ and the Dataloader is eventually terminated with a recursion depth error because _get_bboxes cls_id > self.opt.num_classes is True due to num_classes not being set in opts.py Adding --"num_classes 1" to mot17_tracking.sh solves this. i.epython train_prediction.py tracking --exp_id mot17_motion_model --dataset mot --dataset_version 17trainval --gpus 0 --num_classes 1

[MedChaabane]

thanks for your comments. we will check if num_classes 1 is required in MOT17 training. But for MOT17 you can just use kalman filter as we found that there is no gain in performance when using lstm instead of kalman on MOT17 validation data.