Performance Evaluation of Neuromorphic Visual Attention Based on Proto-Objects

Human beings detects saliency in a visual scene using different mechanisms experimentally studied over the years, such as the detection of "proto-objects", i.e. salient regions in the visual scene where an object could plausibly be. Proto-object saliency-based models’ ability to be both stable and adaptable to any scenario has caught the eye of researchers from both traditional and neuromorphic computer vision, who have often applied bio-inspired visual attention mechanisms to optimise their model. We introduce protoNVA, a novel neuromorphic model for end-to-end saliency detection in event-based visual scenes, drawing inspiration from perceptual grouping mechanisms. Our simulation results show that this model is both more efficient in terms of memory usage and energy consumption, as well as outperforms the saliency detection realised by the state-of-the-art neuromorphic model for proto-objects detection. In addition, protoNVA achieves a better trade-off between architecture weight and saliency detection performance than existing neuromorphic models for visual attention, farther from biology and more suitable for hardware implementation.