Evaluation of Object Tracking Performance of ADNet Method with Different Datasets and Color Spaces
AbstractRecently, object tracking methods based on deep learning have shown great successes. Using deep neural network methods allows following the object even in highly complex scenarios by learning more details of the object in the video and the object motion model. The developed network architecture needs to be trained with a powerful dataset. The variety and size of the selected dataset affect the success of object tracking methods’ results directly. Some object tracking studies have also used image datasets to obtain the diversity in the training set in addition to the video dataset. The object in the image has gained some kind of artificial motion/action between the two images by making the certain rotational or translational movements. But the uses of this technique brings out the question of whether the object has gained a right action or not. In this study, a new perspective is introduced to provide the data the diversity which is required by deep learning-based methods. In this paper, ADnet, which is a deep learning based object tracking method, is used to test our new perspective which is mentioned above. For the training of the ADNet method, color spaces such as HSV, L*a*b*, NTSC, YCbCr, inverted HSV (converted to HSV channeled by taking BGR instead of RGB) were analyzed. As a result of the analysis, HSV and inverted HSV (IHSV) color spaces have been found out to provide stronger and more varied training dataset. Moreover, in order to prepare a proper training set for the object tracking, two different datasets were created with the videos taken from Vot2015 and Vot2014 datasets. Six different training sets were created for each group by translating them into RGB, HSV and IHSV color spaces. A separate ADNet network architecture was trained for each training set. Tests were carried out for each method with the dataset, which included 61 videos. This test dataset includes some videos which are not used in training dataset. In conclusion, it was found that stronger training sets could be created by applying different color transformations such as HSV and IHSV to strengthen the training set.