Efficient Models for Real-time Person Segmentation on Mobile Phones

Despite constantly evolving mobile hardware, real-time person segmentation on mobile phones is challenging due to the limited computational resources. To address this problem, we introduce a novel UNet-like network architecture based on MobileNetV3, which enables the segmentation of persons in images and videos on mobile phones. Our model, which is not limited to a specific shot type, outperforms specialized models in their respective domains and runs with 35 fps on a Google Pixel 4 mobile phone. Moreover, we demonstrate how the segmentation accuracy can be further improved by exploiting the temporal coherence of consecutive frames in videos.