Estimating the correct number of objects in a given natural scene is a common challenge in computer vision. Natural scenes usually contain multiple object categories and varying object densities. Detection-based algorithms are well suited for class-aware object counting and low object counts. However, they underperform with high or varying numbers of objects. To address this challenge, we propose an end-to-end approach to enhance an existing detection based method with a multi-class density estimation branch. The results of both branches are fed into a successive count estimation network, which estimates object counts for each category. Although these numbers do not contain any 10-calization information, they can be used as a valuable indicator for verifying the exactness of the object detector results and improving its counting performance. In order to demonstrate the effectiveness, we evaluate our method on common object detection datasets.
