Robust Discriminative Object Tracker Optimized with KL Divergence

Object tracking is to evaluate the state of a target object in a video sequence at each frame. Despite the astonishing improvements in visual tracking over the years, trackers remain vulnerable to target drift in target estimation when confronting similar distractors or intricate backgrounds. To address these problems, We disaggregate the object tracking task into a conjunction of a classification task and a regression task. For the classification task, Our online multi-layer fusion branch makes the most of target and contextual information. For the regression branch, we adopt the IoU-based method optimized with Kullback-Leibler divergence using Monte Carlo sampling, which models label noise stemming from inaccurate annotations. This simple and effective tracker is termed as MInet. Without bells and whistles, MInet enables leading performance with realtime speeds on five tracking benchmarks comprising OTB2015, VOT2018, LaSOT, UAV123 and GOT-10K, achieving an NP of 0.706 on LaSOT and AO of 0.656 on GOT-10K.