Efficient Joint-Dimensional Search with Solution Space Regularization for Real-Time Semantic Segmentation

Semantic segmentation is a popular research topic in computer vision, and many efforts have been made on it with impressive results. In this paper, we intend to search an optimal network structure that can run real-time for problem. Towards goal, jointly the depth, channel, dilation rate feature spatial resolution, which results space consisting of about $$2.78\times 10^{324}$$ possible choices. To handle such large space, leverage differential architecture methods. However, parameters searched using existing methods need be discretized, causes discretization gap between found by their discretized version as final solution search. Hence, relieve problem from innovative perspective regularization. Specifically, novel Solution Space Regularization (SSR) loss first proposed effectively encourage supernet converge its discrete one. Then, new Hierarchical Progressive Shrinking method presented further achieve high efficiency searching. addition, theoretically show optimization SSR equivalent $$L_{0}$$ -norm regularization, accounts improved search-evaluation gap. Comprehensive experiments scheme efficiently find yields extremely fast speed (175 FPS) small model size (1 M) while maintaining comparable accuracy.