A Normalized Gaussian Wasserstein Distance for Tiny Object Detection

26 Oct 2021  ยท  Jinwang Wang, Chang Xu, Wen Yang, Lei Yu ยท

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

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Object Detection AI-TOD DetectoRS + NWD (ResNet-50-FPN) AP 20.8 # 1
AP50 49.3 # 1
AP75 14.3 # 1
APvt 6.4 # 1
APt 19.7 # 1
APs 29.6 # 1
APm 38.3 # 1
Object Detection VisDrone-DET2019 Cascade R-CNN + NWD AP50 40.3 # 3
APvt 2.9 # 1
APt 11.1 # 1
APs 22.2 # 2

Methods


No methods listed for this paper. Add relevant methods here