bas-22/applied_dl

This paper evaluates the performance of a semi-supervised learning framework, with a Mean Teacher algorithm, used for semantic segmentation of pet images from Oxford-IIIT Pet dataset. In particular, we compare the framework results with a baseline without unlabelled data, and we also compare it with an upperbound with all labels being available. Furthermore, we tackle two research questions. The first question relates to the impact of different noise levels, added to the inputs, on the performance of the framework. The second question focuses on the impact of different labelled data ratios on the framework’s performance. Our experiments show that the framework performs better than the baseline by achieving 78.5% test IoU accuracy which is significantly higher than the baseline which gives 48.9%. The framework with an upper-bound performs the best as it gives 91.2% accuracy. The result of our research on the noise level shows that initially increasing the noise level improves test accuracy, but after a certain level it causes the accuracy to decrease. Lastly, we observe that increasing the labelled that ratio improves model accuracy only to a certain point, suggesting that not having enough labelled data is not the only factor which limits model accuracy, but other factors also play a part.