leonericsson / tracker

Visual Object Trackers on the OTB-mini dataset.

Visual Object Trackers

Visual Object Trackers using the OTB-mini dataset. There are four available trackers: Normalized Cross-correlation, MOSSE Greyscale, MOSSE RGB and finally MOSSE Deep. All trackers were built following tracking theory from Michael Felsberg. The MOSSE trackers have the same foundation but use different features for their discriminative filter where MOSSE Deep uses ResNet50 as a feature backbone.

Dataset

The dataset is a subset of the OTB dataset which can be found here

Evaluation

We mainly use the Success Rate and AUC metrics. These metrics are related, and are both based on the per-image IoU scores for the boxes. Code to produce these scores is given in python_tracker/cvl/dataset.py and an example can be seen in python_tracker/example_tracker.py.

IoU

The IoU is defined per-image, as the area of intersection of the predicted and ground truth bounding box, divided by the area of the union. See python_tracker/cvl/dataset.py for more details.

Success Rate

For any given overlap threshold between [0,1], we can calculate the success rate as the number of frames that achieve a higher IoU than the threshold, divided by the total number of frames. See python_tracker/cvl/dataset.py for more details.

Area Under Curve (AUC)

AUC is a general performance measure which integrates an evaluation metric. In our case we are interested in the mean success rate over the possible overlap thresholds in the success rate metric. See python_tracker/cvl/dataset.py for more details.