Hardware-Aware Affordance Detection for Application in Portable Embedded Systems

Affordance detection in computer vision allows segmenting an object into parts according to functions that those afford. Most solutions for affordance are developed robotics using deep learning architectures require substantial computing power. Therefore, these approaches not convenient application embedded systems with limited resources. For instance, is used smart prosthetic limbs, and this context, could be employed determine the graspable segments of object, which a critical information selecting grasping strategy. This work proposes strategy based on hardware-aware solutions. Experimental results confirmed proposed solution achieves comparable accuracy respect state-of-the-art approaches. In addition, model was implemented real-time devices obtaining high FPS rate, power consumption. Finally, experimental assessment realistic conditions demonstrated method robust reliable. As major outcome, paper characterizes first complete devices. Such substantially improve prosthesis control but it also highly relevant other applications (e.g., resource-constrained robotic systems).