Presentation

These pages describe a dataset for evaluating single-object video tracking algorithms, a corpus of video sequences and ground-truth annotations created to provide a representative benchmark for comparing the performance of tracking algorithms.

The dataset is composed of sequences that cover the most common problems in video tracking (occlusions, illumination changes, similar objects in background, scale changes,...) under different testing conditions (synthetic and real) and type of targets (synthetic ellipses, cars, faces and people).  Moreover, relevant criteria are defined to estimate the complexity for each tracking problem.

For more details please refer to the “Content” section. Recall that this ground-truth is available just for research purposes. 

Additionally, a MATLAB implementation of the evaluation methodology and the SFWDA measure proposed in [1] is available (that defines four performance criteria: global accuracy, temporal efficiency, parameter stability and robustness to target initialization). Please refer to the "dataset downloading" section for a detailed description of the software package.

Main references:

[1]  J.C. SanMiguel, M. Lozano and J. M. Martínez, “Performance evaluation of single object visual tracking: methodology, dataset and experiments”, submitted to (under review), 2013.

[2] M. Lozano Cruz, "Comparative evaluation of video object tracking algorithms", MSc Thesis, February 2012.