Andrija and Jelena: Image Segmentation (PSI:ML 2018)
During our stay in Petnica, we worked on a project called Image Segmentation. We implemented semantic image segmentation in order to recognize different body parts and clothes on “Look into Person” dataset.
We chose this topic because semantic image segmentation is a very present topic in Computer Vision nowadays and can be applied in different situations, which makes it very compelling and relevant.
Our initial task was to label every pixel of an image with one of the 18 different labels. We used a pre-trained convolutional neural network Deep Lab V3 and adapted it for the problem at hand. We also applied changes to its structure in order to improve the network’s performance by replacing the bilinear up-sampling layer with a deconvolutional one. In all cases, we achieved pixel accuracy of above 85%.
During the course of our project, we encountered two main challenges: understanding the role that different layers of the convolutional neural network have in image segmentation and dealing with the large amount of data that was used for training.
We learned how to implement a convolutional neural network in Keras and how to train it for a specific application. We also gained a deeper understanding of how these networks work and which layers are responsible for particular recognition tasks.