Detecting tiny objects in aerial images: A normalized Wasserstein distance and a new benchmark

Tiny object detection (TOD) in aerial images is challenging since a tiny only contains few pixels. State-of-the-art detectors do not provide satisfactory results on objects due to the lack of supervision from discriminative features. Our key observation that Intersection over Union (IoU) metric and its extensions are very sensitive location deviation objects, which drastically deteriorates quality label assignment when used anchor-based detectors. To tackle this problem, we propose new evaluation dubbed Normalized Wasserstein Distance (NWD) RanKing-based Assigning (RKA) strategy for detection. The proposed NWD-RKA can be easily embedded into all kinds replace standard IoU threshold-based one, significantly improving providing sufficient information network training. Tested four datasets, consistently improve performance by large margin. Besides, observing prominent noisy labels Object Detection Aerial Images (AI-TOD) dataset, motivated meticulously relabel it release AI-TOD-v2 corresponding benchmark. In AI-TOD-v2, missing annotation error problems considerably mitigated, facilitating more reliable training validation processes. Embedding DetectoRS, achieves 4.3 AP points improvement state-of-the-art competitors AI-TOD-v2. Datasets, codes, visualizations available at: https://chasel-tsui.github.io/AI-TOD-v2/