Dilated-Scale-Aware Category-Attention ConvNet for Multi-Class Object Counting

Object counting aims to estimate the number of objects in images. The leading approaches focus on single category task and achieve impressive performance. Note that there are multiple categories real scenes. Multi-class object expands scope application task. multi-target detection can multi-class some scenarios. However, it requires dataset annotated with bounding boxes. Compared point annotations mainstream issues, coordinate box-level more difficult obtain. In this paper, we propose a simple yet efficient network based point-level annotations. Specifically, first change traditional output channel from one multiclass counting. Since all use same feature extractor our proposed framework, their features will interfere mutually shared space. We further design multi-mask structure suppress harmful interaction among objects. Extensive experiments challenging benchmarks illustrate method achieves state-of-the-art