Adaptive Linear Span Network for Object Skeleton Detection

Conventional networks for object skeleton detection are usually hand-crafted. Although effective, they require intensive priori knowledge to configure representative features objects in different scale granularity.In this paper, we propose adaptive linear span network (AdaLSN), driven by neural architecture search (NAS), automatically and integrate scale-aware detection. AdaLSN is formulated with the theory of span, which provides one earliest explanations multi-scale deep feature fusion. materialized defining a mixed unit-pyramid space, goes beyond many existing spaces using unit-level or pyramid-level features.Within apply genetic jointly optimize operations connections space expansion. substantiates its versatility achieving significantly higher accuracy latency trade-off compared state-of-the-arts. It also demonstrates general applicability image-to-mask tasks such as edge road extraction. Code available at \href{https://github.com/sunsmarterjie/SDL-Skeleton}{\color{magenta}github.com/sunsmarterjie/SDL-Skeleton}.