A Normalized Gaussian Wasserstein Distance for Tiny Object Detection

• DOI: 10.48550/arxiv.2110.13389
• Publication Date: 2021-01-01
• Authors: Jinwang Wang; Chang Xu; Wen Yang; Lei Yu
• Keywords: Similarity (geometry); Bounding overwatch

Detecting tiny objects is a very challenging problem since object only contains few pixels in size. We demonstrate that state-of-the-art detectors do not produce satisfactory results on due to the lack of appearance information. Our key observation Intersection over Union (IoU) based metrics such as IoU itself and its extensions are sensitive location deviation objects, drastically deteriorate detection performance when used anchor-based detectors. To alleviate this, we propose new evaluation metric using Wasserstein distance for detection. Specifically, first model bounding boxes 2D Gaussian distributions then dubbed Normalized Distance (NWD) compute similarity between them by their corresponding distributions. The proposed NWD can be easily embedded into assignment, non-maximum suppression, loss function any detector replace commonly metric. evaluate our dataset (AI-TOD) which average size much smaller than existing datasets. Extensive experiments show that, equipped with metric, approach yields 6.7 AP points higher standard fine-tuning baseline, 6.0 competitors. Codes available at: https://github.com/jwwangchn/NWD.