Semi-supervised Semantic Segmentation using Generative Adversarial Networks
Author
Jensen, Troels Høg Peter
Term
4. term
Publication year
2017
Abstract
This master's thesis investigates whether generative adversarial networks (GANs) can support semi-supervised semantic segmentation for robot perception, aiming to reduce reliance on fully labeled data. The proposed system encourages a generator to produce segmentation-like masks and integrates a Region Proposal Network (RPN) to supply candidate regions of interest to the model. Positioned as an alternative to conventional supervised convolutional neural networks, the work outlines the overall design, implementation, and training and processing components. The results indicate that the GAN-based approach did not yield satisfactory segmentation masks and was not adequate for practical use; the thesis discusses these limitations and suggests directions for future work in semi-supervised segmentation.
Denne kandidatafhandling undersøger, om generative adversarial networks (GAN'er) kan anvendes til semi-superviseret semantisk segmentering i robotperception for at reducere behovet for fuldt annoterede data. Det foreslåede system opmuntrer en generator til at producere maskebilleder, der ligner segmenteringsresultater, og integrerer en Region Proposal Network (RPN), som leverer kandidatområder af interesse til modellen. Tilgangen præsenteres som et alternativ til klassisk, fuldt superviseret segmentering med konvolutionsnetværk, og arbejdet beskriver overordnet design, implementering samt trænings- og behandlingskomponenter. Resultaterne viser, at den GAN-baserede metode ikke opnåede tilfredsstillende segmenteringsmasker og derfor ikke levede op til praktiske krav; afhandlingen diskuterer disse begrænsninger og peger på mulige retninger for fremtidig forskning i semi-superviseret segmentering.
[This apstract has been generated with the help of AI directly from the project full text]