Article directory
Summary
Detecting tiny objects is an extremely challenging problem because tiny objects only contain a few pixels in size. Due to the lack of appearance information, state-of-the-art object detectors cannot produce satisfactory results on tiny objects. Because IoU-based metrics such as IoU (intersection-over-union ratio) and their extensions are very sensitive to position deviations of small objects, and will drastically deteriorate detection performance when used in anchor-based detectors. To alleviate this problem, the authors propose a new evaluation metric for small object detection, namely using Wasserstein distance. The bounding boxes are first modeled as a 2D Gaussian distribution, and then a new metric called Normalized Wasserstein Distance (NWD) is proposed to calculate their similarity through the corresponding Gaussian distribution. The proposed NWD metric can be easily embedded into the assignment, non-maximum suppression and loss functions of any anchor-based detector to replace the commonly used IoU metric. The authors evaluate NWD on a new Tiny Object Detection (AI-TOD) dataset, where the average object size is much smaller than existing object detection datasets. Experiments show that