Fast Portrait Segmentation of the Head and Upper Body

Portrait segmentation is the process whereby the head and upper body of a person is separated from the background of an image or video stream. This is difficult to achieve accurately, although good results have been obtained with deep learning methods which cope well with occlusion, pose and illumination changes. These are however, either slow or require a powerful system to operate in real-time. We present a new method of portrait segmentation called FaceSeg which uses fast DBSCAN clustering combined with smart face tracking that can replicate the benefits and accuracy of deep learning methods at a much faster speed. In a direct comparison using a standard testing suite, our method achieved a segmentation speed of 150 fps for a 640x480 video stream with median accuracy and F1 scores of 99.96% and 99.93% respectively on simple backgrounds, with 98.81% and 98.13% on complex backgrounds. The state-of-art deep learning based FastPortrait / Mobile Neural Network method achieved 15 fps with 99.95% accuracy and 99.91% F1 score on simple backgrounds, and 99.01% accuracy and 98.43 F1 score on complex backgrounds. An efficacy-boosted implementation for FaceSeg can achieve 75 fps with 99.23% accuracy and 98.79% F1 score on complex backgrounds.