Deep Learning (DL) and especially Convolutional Neural Network (CNN) has drawn a lot of attention in last few years. At present day CNN has been successfully applied in various fields of Computer Vision, for example: image classification [1, 2, 3], semantic segmentation [4], object detection [5], human pose estimation [6], etc. Also some of the recent researches have tried to apply discriminating power of CNN for visual tracking purposes: [7, 8, 9, 10, 11]. Almost all of them outperform state-of-art tracking algorithms (like TLD [12], KCF [13]) in terms of accuracy, but lose in terms of speed. One of the most recent successful trackers based on CNN are MDNet and GOTURN. Though MDNet has been evaluated as a top – accuracy tracker (1-st place on VOT2015), it suffers from low performance (~1 FPS) due to CNN online learning. While GOTURN can run on around 100 FPS maintaining high accuracy and robustness (close to MDNet). Good speed/accuracy tradeoff makes GOTURN one of the most promising next-generation trackers. So this proposal aims to implement GOTURN tracker in OpenCV library.



Tyan Vladimir


  • Antonella Cascitelli