This paper addresses the question of whether edge information enables tracking objects reliably. Human visual system is adept at tracking shapes without any texture. Motivated by this, we incorporated an object proposal mechanism that uses sparse yet informative contours to score proposals based on the number of contours they wholly enclose into a detection-by-tracking process for visual tracking. Our method is able to execute search in the entire image quickly and focus only on those high-quality candidates to test and update our discriminative classifier. Using high-quality candidates to chose better positive and negative samples, we reduce the spurious false positives and improve the tracking accuracy. Since our tracker employs only a few candidates to search the object, it has potential to use higher-dimensional features if needed. More importantly, our method can track randomly and very fast moving objects. It is robust to full occlusions as it is able to rediscover the object after occlusion. The presented tracker outperforms all the state-of-the-art methods on four largest benchmark datasets.
View on arXiv