Deeply scale aggregation network for object counting

摘要

Object counting is a fundamental and essential task to build the high-level cognition for the given scene in the computer vision field. For the common scale variant issue in the object counting, this paper designs a deeply scale aggregation network for object counting task. Specially, we design a dense multi-scale (DMS) block to extract the initial scale-aware feature, and thus stack multiple DMS blocks in the specific structure to implement the adaptive network depth, so as to learn the more mighty scale-aware feature. Extensive experimental results are reported in this paper, using up to six public object counting benchmarks, which demonstrate that the proposed method has an effective performance for object counting task. And the ablation studies validate the structural rationality of the proposed method.