search

SilvioGiancola / TrackingNet-devkit

Development kit for TrackingNet
https://tracking-net.org/
170 stars 32 forks source link

How can I exclude invalid annotation frames? #24

Open zosel260 opened 3 years ago

zosel260 commented 3 years ago

Thanks you for sharing good dataset.

I downloaded TRAN_0 set and annotations from https://exrcsdrive.kaust.edu.sa/exrcsdrive/index.php/s/MAaiTPdOwiPDNlp

I fond out some sequences have incorrect ground truth rect information. For example , TRAIN_0/ a1qoB1eERn0_0 : around the frame number 155-170, gt rects are incorrect. TRAIN_0/ a3In51YCqMg_0 : around the frame number 457-477, gt rects are incorrect.

I'd like to exclude these incorrect frames to train my model. Is there any annotation(indicator) for these frames?

Thanks in advance.

SilvioGiancola commented 3 years ago

As mentioned in our paper, the dense annotations are extracted from coarser annotations at 1fps using a mixture of trackers.

Feel free to use any heuristic that discard those annotations if they are not suitable for your training.