Real-time segmentaion of feet on smartphone

Sammanfattning: Image segmentation, the problem of dividing an image into meaningful parts and semantic segmentation, dividing images into multiple known classes are very interesting problems in computer vision. Segmentation is relevant for everything from self-driving cars to 3D-modeling and synthetic camera effects. The current state of the art inthese problems is achieved with computationally intensive deep convolutional neural networks. Even though there currently exist modelsfor producing high-quality segmentations, most of these are too resource intensive to run in real-time on mobile devices, something that is more and more requested for user experience in computer vision apps and safety in self-driving cars. To build models capable of real-time segmentation on smartphones this thesis builds upon previous work on constructing streamlined neural networks and compression of these using effective approximations of the convolutional operator and student-teacher training techniques. Specifically, this is done for segmentation of feet from the background which is a part of a larger project for 3D-scanning of feet on smartphones conducted at Volumental. The fastest neural network produced in this work manages to runat over 10f ps on an Android smartphone from 2016 without the use of hardware acceleration. Despite this speed, the network manages to achieve a mean intersect over union (mIoU ) score of over 93% on a held out test set.