Evaluation Metrics and Toolkit

In LaSOT, we conduct one-pass evaluation (OPE) to assess the performance of each tracker. In detail, we utilize three types of metrics, Precision, Normalized Precision and Success, to measure different tracking algorithms. The definitions of the three metrics can be seen in the paper.

The evaluation toolkit can be found in Github, or download the complete package with tracking results inside here.

Evaluation Protocol

We define two protocols for evaluating trackers on LaSOT as follows

  • Protocol I: All 1,400 sequences in LaSOT are employed for evaluation. Researchers are allowed to leverage any videos except for those in LaSOT for training their trackers.
  • Protocol II: Only the testing subset (280 sequences) of LaSOT is utilized for evaluation. Research are only allowed to leverage the training set of LaSOT to develop their trackers. (Training/Testing split: Training Subset | Testing Subset)

Evaluated Trackers

We assess 35 popular tracking algorithms on LaSOT under Protocol I and II. These trackers include deep learning based ones, correlation filter based ones with hand-crafted or deep features, sparse representation based ones and other representatives. Table 1 shows these trackers.

Table 1. Description of each tracking algorithm in detail.
Tracker Paper Where When Speed Code

Note: Each tracker is used as it is from authors' implementation, without any modification.

Evaluation Results

The following plots demonstrate the evaluation results of 35 trackers under two protocols using three metrics.

Plots of Protocol I

Plots of Protocol II