Visual Content-Agnostic Novelty Detection Engine with 2.4 pJ/pixel Energy and Two-Order of Magnitude DNN Activity Reduction in 40 nm

Abstract

An engine to identify frames with novel content in a video stream is proposed as additional vision pipeline stage following conventional saliency detection. Based on connected component analysis with mean-center tracking, its complexity is reduced to linear compared to quadratic in prior art. This introduces frame-level temporal sparsity for subsequent DNN activity/power reduction (177X beyond saliency detection). 2.4 pJ/pixel energy is achieved in 40 nm.