Small Object Detection via Scale-Adaptive Label Assignment and Localization Uncertainty

Despite the current detectors achieving outstanding performance, detecting small objects remains a challenging problem. The challenge mainly arises from the low quantity and quality of samples as well as the inherent difficulty in localization. Focusing on these problems, we present an approach for small object detection with a scale-adaptive label assignment scheme and a novel quality-driven localization loss (QLL). First, we perform the scale-adaptive label assignment by combining distance-based and Intersection-over-Union (IoU)-based criterion along with a scale discriminator mechanism to obtain larger quantity and higher quality of training samples. Then, we extend an additional branch parallel to the original localization branch, modeling the localization task as predicting Gaussian probability distributions to incorporate localization uncertainty. Finally, we develop QLL by integrating the scale information and IoU to achieve more accurate localization for small objects. Extensive experiment results on two natural images benchmarks demonstrate that our method underscores its superiority over baseline detector in detecting small objects. Moreover, our method performs better than other recent state-of-the-art methods on the large-scale small object detection benchmark SODA-D without bells and whistles.