Beyond semi-supervised tracking: Tracking should be as simple as detection, but not simpler than recognition

Abstract

We present a multiple classifier system for model-free tracking. The tasks of detection (finding the object of interest), recognition (distinguishing similar objects in a scene), and tracking (retrieving the object to be tracked) are split into separate classifiers in the spirit of simplifying each classification task. The supervised and semi-supervised classifiers are carefully trained on-line in order to increase adaptivity while limiting accumulation of errors, i.e. drifting. In the experiments, we demonstrate real-time tracking on several challenging sequences, including multi-object tracking of faces, humans, and other objects. We outperform other on-line tracking methods especially in case of occlusions and presence of similar objects.

Visual tracking via weakly supervised learning from multiple imperfect oracles

“…For performance evaluation, we compare our approach against several representatives of the current state-of-the-art in visual tracking -the Fragments-based Tracker [5], the Online Boosting Tracker [3], the Semi-Supervised Online Boosting Tracker [8], the Beyond Semi-Supervised Tracker [9], the Online Multiple Instance Learning-based Tracker [10], and the SURF Tracker [16]. In the rest of our experiments, we refer to these six compared algorithms as FT, OBT, SSOBT, BSST, OMILT and ST respectively.…”

“…But the prior might be too strong (i.e., limited appearance changes and partial occlusions) and generic (i.e., no discrimination between different objects from one class). (4) Beyond Semi-Supervised Tracker [9]. The tracker uses beyond semi-supervised tracking method, which is balancing between semi-supervised and the fully adaptive tracking, to obtain solutions.…”

“…(1) The set consists of a variety of complementary tracking approaches, such as gray histogram-based patch matching [5], motion consensus of local descriptors [16] and online classification using haar feature [3,8,9,10]. (2) The features used in all the six tracking methods can be extracted in a very efficient manner due to integral image data structure.…”

“…Despite their success, these approaches are limited by the fact that they cannot accommodate very large changes in appearance. To balance between semi-supervised and the fully adaptive tracking, Stalder et al [9] present an approach using object specific and adaptive priors. In [10], Babenko et al propose to use a multiple instance learning based appearance model for object tracking.…”

Hierarchy of Localized Random Forests for Video Annotation

“…Usually a bounding box is propagated to later frames. Boosting based classifiers [17] and random forest classifiers [16] are popular choices for learning and updating the template model. Godec et al [10] also give a rough segmentation of the object being tracked.…”