Comparison and Optimization of U-Net and SegNet Encoder-Decoder Architectures for Soccer Field Segmentation in RoboCup

Deep Neural Networks are considered state-of-the-art for computer vision tasks. In the humanoid league of the RoboCup competition, many teams have relied on neural networks for their computer vision systems, especially after the rules were changed to be closer to the ones used in human soccer. One of the main vision tasks solved using neural networks in this domain is soccer field segmentation, where an algorithm must classify each image pixel. This task has been solved classically with simple color segmentation, but recently, the teams have been migrating to encoder-decoder convolutional neural networks. The segmented image is then post-processed by another algorithm that extracts information about field features such as the lines and the field boundary. In this article, the contribution is a comprehensive comparison regarding how different neural networks perform in the soccer field segmentation task, considering the constraints imposed by RoboCup. Twenty-four neural network models, based on the U-Net and SegNet architectures, were tested and compared in terms of loss, Dice Scores, and average inference time to examine various encoders’ efficiency. The U-Net/MobileNetV3Small and U-Net/MobileNetV3Large models, particularly, achieved Dice Scores of 0.7173 and 0.7107 for line segmentation, respectively. Furthermore, to make the models suitable for real-time inference on the embedded computer of a small humanoid robot, we optimized the models for faster inference on the CPU using the Intel OpenVINO framework, with the U-Net/MobileNetV3Large showing a Dice Score of 0.733 after optimization.