Model initialisation is an important problem in object tracking. Tracking algorithms are generally provided with the first frame of a sequence and a bounding box indicating the location of the object. This bounding box may contain of a large number of background pixels in addition to the object and can lead to parts-based tracking algorithms initialising their object models in background regions of the bounding box.
We tackled this as a missing labels problem, marking pixels sufficiently away from the bounding box as belonging to the background and learning the labels of the unknown pixels. We adapted three techniques to this problem; One-Class SVM (OC-SVM), Sampled-Based Background Model (SBBM), a novel background model based on pixels samples, and Learning Based Digital Matting (LBDM).
These were evaluated with leave-one-video-out cross-validation on images taken from the VOT2016 tracking benchmark. Our evaluation showed both OC-SVMs and SBBMs are capable of providing a good levels of segmentation accuracy but are too parameter-dependent to be used in real-world scenarios. We showed that LBDM achieved significantly increased performance with cross validation selected parameters and investigated its robustness to parameter variation.
This repository contains the three Python 3.6 implementations of the techniques (LBDM, SBBM, OC-SVM) used in the paper.
- All techniques: matplotlib, numpy, scipy, skimage
- OC-SVM: sklearn
Information on each method can be found by typing help(function_name)
at the python interpreter.
To see a working example of each of the three techniques, run their respective python files:
python alpha_matting_segmentation.py
python ocsvm_segmentation.py
python sbbm_segmentation.py
This project is licensed under the MIT Licence - see the LICENCE file for details.