Human instance segmentation from video using locally competing 1SVMs with shape prior

Abstract

In this paper, we propose a method for human segmentation in videos, extending the recent locally competing 1SVM model. There are only local color distributions to be made use of in the model. To generate a consistent segmentation from complex environments, first, we assume we obtain a bounding box around human by using the human detector. Then we incorporate shape prior information inside the bounding box, which biases the segmentation towards typical human shapes. Finally, we show a substantial improvement over C-1SVM method from our experiment.