Simone Mentasti and Matteo Matteucci
Politecnico di Milano, Italy
In this paper we describe a lightweight framework for fast image segmentation on embedded systems, based on superpixels, which leverages on convolutional and graph-convolutional neural networks. In particular, we analyzed different superpixel representation looking for the best trade-off between the efficiency of the system and richness of the description. Similarly, we analyzed different network sizes, balancing the number of filters used and the prediction accuracy. We also compared two different convolutional architecture, one based on the classical encoder-decoder paradigm and one based on graphs, to guarantee a most accurate representation of the image structure. The architecture was tested on the KITTI dataset using an embedded system with CUDA capabilities.
