Fully convolutional networks for segmenting images from an embedded camera

This paper describes a Fully Convolutional Network (FCN) to segment images from a compact stereo imaging sensor attached to a robot, i.e., a configurable DUO M® stereo camera embedded onto a robot to provide low-level computer vision functions. This robot is named LEIA-1 and carries a NVIDIA Jetson TK1® platform to execute high-level robotics vision, planning and decision making algorithms. Since the TK1 board is not dedicated solely to robotics vision, the FCN architecture should be light enough to consume limited resources. The FCN described in this paper was implemented and prototyped with Python, Keras, and Theano. Also, we trained and validated the FCN architecture using an adaptation of the dataset known as Playing for Data to match the embedded camera specifications. The results reveal the viability of integrating FCNs to processing platforms attached to robots and use their SoC/GPU power to segment indoors/outdoors captured images.