单目标跟踪数据集

LaSOT . LaSOT has 1,400 sequences in 70 categories, which amounts to a total of 3.52M frames. Each category contains exactly twenty sequences, making the dataset balanced across classes. It also provides longer sequences that contain more than 1,000 frames (2,512 frames on average) in order to satisfy the current long-term trend in tracking.LaSOT dataset follows the OPE criterion of OTB . It consists of precision plot which is measured by the center location error, and success plot which is measured through the intersection over union (IoU) between the predicted bounding box and the ground-truth. Besides precision plot and success plot, LaSOT also uses normalized precision plot to counter the situation that target size and image resolution have large discrepancies for different frames and videos, which heavily influences the precision metric.

VOT2018/16 . VOT2018 dataset has 60 public testing sequences, with a total of 21,356 frames. It is used as the most recent edition of the VOT challenge. The VOT protocol establishes that when the evaluated tracker fails, i.e. when the overlap with the ground-truth is below a given threshold, it is re-initialized in the correct location five frames after the failure. The main evaluation measure used to rank the trackers is Expected Average Overlap (EAO), which is a combination of accuracy ( A ) and robustness ( R ). We also use VOT2016 for comparison purposes, which has 10 different sequences with VOT2018.

TrackingNet . This is a large-scale tracking dataset consisting of videos in the wild. It has a total of 30,643 videos split into 30,132 training videos and 511 testing videos, with an average of 470,9 frames. It uses precision, normalized precision and success as evaluation metrics.

之后再继续总结…