A Double Diagonal Ratio Metric for Tiny Object Detection in Aerial Images
Haoguang Liu, Qiang Tong, Lin Qi, Yanyan Hao, Xiulei Liu · 2023
In recent years, with the development of deep learning and the maturity of object detection technology, unprecedented progress has been made in aerial image object detection tasks. But for detecting tiny objects with little pixel information and easily confused with the background, the general-purpose detectors still do not achieve good performance. Through analysis, we found that, in addition to the defects of the tiny objects themselves, the sensitivity of the most commonly used IoU-based metric to tiny objects is one of the main reasons for the poor detection performance. Therefore, we design a new metric, Double Diagonal Ratio (DDR), which alleviates the poor performance by using a first power line segment ratio instead of a second power area ratio. Note that, DDR is simple in design and can be easily embedded into almost all detectors assignment, non-maximal suppression, and loss function modules to replace IoU. In addition, DDR does not introduce additional hyper-parameters and has stronger robustness. We evaluate our metric on the AI-TOD dataset, and extensive experimental results show that DDR significantly affects the detector performance improvement, especially on tiny objects.