Rethinking Rotated Object Detection with Gaussian Wasserstein Distance Loss

By Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, Qi Tian. International Conference on Machine Learning, 2021

609 citations recorded by the source corpus.

Boundary discontinuity and its inconsistency to the final detection metric have been the bottleneck for rotating detection regression loss design. In this paper, we propose a novel regression loss based on Gaussian Wasserstein distance as a fundamental approach to solve the problem. Specifically, the rotated bounding box is converted to a 2-D Gaussian distribution, which enables to approximate the indifferentiable rotational IoU induced loss by the Gaussian Wasserstein distance (GWD) which can be learned efficiently by gradient back-propagation. GWD can still be informative for learning even there is no overlapping between two rotating bounding boxes which is often the case for small object detection. Thanks to its three unique properties, GWD can also elegantly solve the boundary discontinuity and square-like problem regardless how the bounding box is defined. Experiments on five datasets using different detectors show the effectiveness of our approach. Codes are available at https://github.com/yangxue0827/RotationDetection and https://github.com/open-mmlab/mmrotate.

Subfields: Artificial Intelligence, Biomedical Engineering, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/b010c4a15eb6e97a7410c3eff4520075d93d368c

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Wenwei Zhang
2. Junchi Yan
3. Qi Ming
4. Wentong Li
5. Rongguang Ye
6. Xue Yang
7. Qi Tian
8. Jirui Yang

In Biomedical Engineering

1. Qi Tian
2. Kai Chen
3. Wangmeng Zuo
4. Qibin Hou
5. Shilong Zhang
6. Jian Yang
7. Liang Zheng
8. Yi Yang

In Computer Vision and Pattern Recognition

1. Qi Tian
2. Luc Van Gool
3. Junchi Yan
4. Wangmeng Zuo
5. Ming-Hsuan Yang
6. Xue Yang
7. Jianke Zhu
8. Lei Zhang

Citations in the corpus

• Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence (2021, methodology): “We divide the images into 600× 600 subimages with an overlap of 150 pixels and scale it to 800× 800, in line with the cropping protocol in literature [5, 27].”
• Localization Distillation for Object Detection (2022, methodology): “Parametric regression is the most popular manner in the classical dense regression-based rotated object detection [71], [73], [74], [75].”
• RTMDet: An Empirical Study of Designing Real-Time Object Detectors (2022, background): “, Gaussian distribution [89,90] and convex set [26, 44]) to ease the rotated bounding box regression task.”
• Large Selective Kernel Network for Remote Sensing Object Detection (2023, background): “Additionally, several loss functions, including GWD [77], KLD [79], and Modulated Loss [54], have been further proposed to enhance the Figure 1.”
• Detecting tiny objects in aerial images: A normalized Wasserstein distance and a new benchmark (2022, neutral): “The poor performance of GWD and IoU-based metrics (i.e., GIoU, DIoU, IoU) mainly results from the fact that they were originally designed as loss functions without a normalized value range.”
• Oriented RepPoints for Aerial Object Detection (2021, methodology): “Although having achieved promising performance, the direct orientation prediction still has some issues including the discontinuity of loss and regression inconsistency [25, 40, 41, 43].”
• A Normalized Gaussian Wasserstein Distance for Tiny Object Detection (2021, neutral): “[32] also propose a Gaussian Wasserstein Distance (GWD) loss for oriented object detection by measuring the positional relationship of oriented bounding boxes.”
• Poly Kernel Inception Network for Remote Sensing Detection (2024, methodology): “Consequently, numerous prior efforts have been dedicated to developing various oriented bounding box (OBB) detectors [10, 20, 31, 65, 67, 71] and improving the angle prediction accuracy for OBBs [68, 70, 72–74].”
• MMRotate: A Rotated Object Detection Benchmark using PyTorch (2022, methodology): “Depending on the angle range, it can be divided into OpenCV definition ( 𝐷 𝑜𝑐 ,𝜃 ∈ [− 𝜋 / 2 , 0 ) ) [32, 36, 38], long edge 90° definition ( 𝐷 𝑙𝑒 90 ,𝜃 ∈ [− 𝜋 / 2 , 𝜋 / 2 ) ) [3, 6], and long edge 135° ) ) [5]. https://github.com/dingjiansw101/AerialDetection 4 https://github.com/jbwang1997”