pre_seq question

[NJiHyeon]

hello! I think it made the pre_seq_list through explore function, but is it possible to take it through groundtruth coordinates and learn it with past coordinates?

[AlexDotHam]

I think it will be ok, but maybe get the lower precision, we tried to do that in the beginning, but I didn't get better accuracy, maybe I was wrong? :(

[NJiHyeon]

I'm also trying to use groundtruth past coordinates after thinking about many ways. Do I need to modify it in the sequence_sampler.py file to use the above method? May I know how you made it before?

[AlexDotHam]

Without sequence_sampler, but using sampler enough, you can sample the search region and template in a video, after that, get the ground truth before the search frame in 7 frames, and train like one-stage but without any augmentation like random jitter or rescale. In my opinion, that kind of training is too easy for a model to gain generalization.

[NJiHyeon]

"that kind of training is too easy for a model to gain generalization" -> What exactly does this mean? Are you saying that the model does not perform well? And I wonder to what extent the performance has decreased.

[AlexDotHam]

In my opinion, if you give the ground truth of the video, you may have leaked potential sports patterns. This makes the training a mundane task, that will guide the model to try his best to simulate trajectory transformation rather than read the images in the current frame. In our experiments, this processing will lead accuracy drop compared to second-stage, but still better than solely trained in a one-stage.

[NJiHyeon]

Thank you so much for the detailed explanation.