Godec et al. use Hough forests and ferns for online object tracking in images. They implement voting-based detection and back-projection and a rough GrabCut segmentation approach, removing the bounding box representation of the non-rigid objects influenced under transformations, partial occlusions, scale changes, and rotations.

The method compares image pixels to detect non-rigid objects in online learning-based tracking. It uses Hough-based classification to implement randomized Hough ferns for discrete region classification. This technique combines back-projection and GrabCut segmentation for pixel-wise separation of the object from the image background. Hence, this validates the geometric relation to achieve fine-grained tracking-by-detection for non-rigid objects.

The authors derive the HoughTrack algorithm, introducing online Hough forests and ferns as an extension to random forests and ferns. This initiates Hough voting to help decide the object’s center based on vectors of the object’s foreground features. In this process, the center of mass in the foreground region is shifted from the object’s center, reconfiguring the object geometry. The algorithm tackles segmentation failure problems by using bounding box representation as an alternative to wrongly labeled samples.

The experiment presents a clear picture of the online learning-based tracking performed over a bounding box dataset and a set of sequences from part-based tracking approaches. The paper also highlights the space and time complexity of the proposed approach. The authors suggest “using a more efficient segmentation algorithm and parallelizing the fern implementation” to speed up the whole procedure. Online learning is a new idea in object tracking involving non-rigid objects, which makes this paper worthwhile reading for a better understanding of the tracking-by-detection method.