About the evaluation script on the validation set

[davidyang180]

Hello! I saw in the paper that the evaluation index of MOTA is 71.3 reported on the validation set in the ablation experiment. I found that there are only evaluation scripts for the test set and training set in the code. May I ask when I want to evaluate on the validation set, I need to manually select Related video sequences?

[timmeinhardt]

There is no dedicated validation set. We validate on a cross-validation split on the training sequences. The split is a per -sequence frame split. See for example mot17_train_cross_val_frame_0_0_to_0_5_coco in src/generate_coco_from_mot.py. To only evaluate a subset of frames the src/track.py has a frame_range parameter. This means you can run python src/track.py with frame_range.start=0.5.

[davidyang180]

There is no dedicated validation set. We validate on a cross-validation split on the training sequences. The split is a per -sequence frame split. See for example mot17_train_cross_val_frame_0_0_to_0_5_coco in src/generate_coco_from_mot.py. To only evaluate a subset of frames the src/track.py has a frame_range parameter. This means you can run python src/track.py with frame_range.start=0.5.

Thank you for your reply. For the ablation experiments in some MOT papers, half of the training and half are used as the validation set. In your paper, do you use all of them for training, and then use the latter half for the validation of the ablation experiments?

[davidyang180]

There is no dedicated validation set. We validate on a cross-validation split on the training sequences. The split is a per -sequence frame split. See for example mot17_train_cross_val_frame_0_0_to_0_5_coco in src/generate_coco_from_mot.py. To only evaluate a subset of frames the src/track.py has a frame_range parameter. This means you can run python src/track.py with frame_range.start=0.5.

Hi! Another problem is that the paper mentions that the relevant details of the model will be further described in the appendix. Where can I download the appendix of this paper?

[timmeinhardt]

For the ablation experiments, we use do 50-50 frame split on the training sequences. Training on the first 50% of the frames of each training sequences and validating on the second 50%. For our final submission and test set evaluation we train on all frames of the training sequences.

The arxiv paper has the appendix at the end.