Deep learning has basically occupied most of the branches in the field of computer vision around the middle of 2015, such as image classification, object detection, etc., but no visual tracking work has been announced. At the end of 2015, a paper called MDNet appeared, dedicated to For solving visual tracking with neural networks, it was also the 2015 VOT champion.
Convolutional Neural Networks (CNN) have some basic problems:
A necessary condition for CNN to achieve great success in image classification is to have millions of labeled training datasets called ImageNet. For visual tracking, there has been a lack of datasets with a similar amount of data. If only the ground truth given in the first frame of the video is used as training data, it will undoubtedly be very easy to overfit.
