Improving object counting in aerial images with density-based threshold shifting

Abstract

ABSTRACT This paper presents an approach to reduce the negative effects of object density when objects are counted via detection in aerial images. A novel image dataset was generated and used to fine-tune an existing object detection architecture. A novel video dataset was also generated and used for the development and testing of the proposed approach. The datasets consist of aerial images and videos of sheep with scenarios of tightly clustered and isolated individuals. This ensures both sparse and dense object distributions within our dataset. The proposed approach is compared to a proven detection-based counting baseline. In this work, it is explicitly shown that object density and classification probability have a linearly inverse relationship and that reductions in classification probabilities, caused by high densities, has a significantly negative impact on counting performance. The proposed approach uses density-based threshold shifting to improve counting performance, i.e. by dynamically adjusting the counting threshold, based on the density, we are able to improve counting performance. The proposed approach reduces the overall mean absolute error by 78.51% compared to the baseline on an unseen test dataset. It seamlessly integrates with most existing object detection or instance segmentation frameworks without any modifications. The proposed approach also reduces the sensitivity of the counting threshold selection, implying that competitive results can be achieved with minimal tuning. The code is made publicly available. 1