This use case compares formal neural approach with both spiking and hybrid ones for different kind of applications that are compatible with drone missions. It aims to implement several kinds of applications related to images acquired from devices evolving at high altitudes. Such images have in common large resolutions while object of interest can be no wider than few pixels. The foreseen applications cover classification, detection and segmentation and can be used in agriculture applications, wildfire detection and management, border surveillance and others.
TRT, in collaboration with CEA and UZH, implements several kind of missions to reflect the variety of potential applications and evaluates their executions using different type of hardware (formal ASIC, spiking ASIC, formal/hybrid FPGA IP) developed in ANDANTE to identify which neural network fits best on what accelerator in order to significantly reduce the power consumption while maintaining the state-of-the-art accuracy and inference time.