MBest Struct: M-Best diverse sampling for structured tracker

Ivan Bogun and Eraldo Ribeiro

Abstract

We approach the problem of model-free visual tracking of objects in videos. Model-free tracking has its state-of-the-art in a class of methods called tracking-by-detection, as shown in recent benchmarks. Some top-performing methods use deep neural networks (i.e., convnets) to solve the learning-based steps of the tracking algorithm (e.g., bounding-box prediction and evaluation). Despite improving accuracy, convnets impose a high computational cost on trackers, limiting their real-time applications. In this paper, we propose to use deep features from a pre-learned deep-convolutional network in a computationally efficient way. Here, we use M-Best diverse-sampling to sample a small yet diverse set of bounding boxes that are likely to contain the tracked object. Given these bounding boxes, our method performs detection using deep features. The resulting tracker, named MBestStruck, uses a high-quality feature representation while being computationally efficient. Our tracking approach compares very well to the state-of-the-art, as shown by experiments done on popular benchmark datasets.

Session

Posters 2

Files

Extended Abstract (PDF, 242K)
Paper (PDF, 5M)
Supplemental Materials (ZIP, 10M)

DOI

10.5244/C.30.93
https://dx.doi.org/10.5244/C.30.93

Citation

Bibtex

        @inproceedings{BMVC2016_93,
        	title={MBest Struct: M-Best diverse sampling for structured tracker},
        	author={Ivan Bogun and Eraldo Ribeiro},
        	year={2016},
        	month={September},
        	pages={93.1-93.12},
        	articleno={93},
        	numpages={12},
        	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
        	publisher={BMVA Press},
        	editor={Richard C. Wilson, Edwin R. Hancock and William A. P. Smith},
        	doi={10.5244/C.30.93},
        	isbn={1-901725-59-6},
        	url={https://dx.doi.org/10.5244/C.30.93}
        }