Combinatorial progressive architecture search for crowd counting

Abstract

Recently, the automated machine learning system, including neural architecture search (NAS), has been introduced to the task of crowd counting. However, there are several concerns about applying existing AutoML methods to crowd counting: (i) The previous AutoML system automates the CNNs of crowd counting without acceleration design and ignores the reproducibility on edge computation circumstances; (ii) a simple concatenation of the fused multi-scale information cannot reflect the pixel-wise information; thus it demands a high-level combinatorial system to learn these singular pieces of information jointly. In this paper, we propose an efficient automation system to tackle the lightweight neural architecture search and combinatorial problem. At first, we design a combinatorial search space for the automatic combination of the singular information (tensors). We then introduce a progressive neural architecture search mechanism that progressively combines the information to be more simple and precise. For efficiency, the combination controller controls both the depth and width of the search space, acquiring simple shallow combinations first and complex and heavy ones later. Our mobile-based searched feature fusion architecture shows strong transferability with the use of the recurrent neural network as a controller and combinatorial dynamic system. The proposed combinatorial progressive architecture search (CPAS) algorithm achieves competitive results on benchmark datasets and sets the new state-of-the-art results on the UCF-CC 50 dataset. The proposed CPAS system acquires to-edge ability.