Tracker Evaluation for Small Object Tracking

Chang Liu, Chunlei Liu, Linlin Yang, Baochang Zhang

Abstract: The small object problem becomes an increasingly important task because of its wide application. There are three significant challenges for small objects: 1) small objects have extremely vague and variable appearances, 2) due to the low resolution of the input images, their characteristic expression information is inadequate and, therefore, is prone to be absent after downsampling and 3) they draft drastically in the images when lens shake violently. Even though small object detection has been extensively studied, small object tracking is still in its infancy. To further explore small object tracking, we evaluate six latest trackers on OTB100 (normal object dataset) and small90 (small object dataset). According to our observation, we draw three instructive conclusions for the follow-up research of small object tracking. Firstly, due to the weak characteristics of small objects, existing trackers perform worse on small objects than on normal objects. Secondly, based on the results of ATOM, SPSTracker, DIMP, SiamFC and SiamMask, the trackers’ performance on small objects is positively correlated with that on normal objects. Thirdly, trackers tend to perform better on small object datasets when they can handle drift, occlusion and out-of-view.