Learning Modulated Loss for Rotated Object Detection

By Wen Qian, Xue Yang, Silong Peng, Yue Guo, Chi Yan. AAAI Conference on Artificial Intelligence, 2019

399 citations recorded by the source corpus.

Popular rotated detection methods usually use five parameters (coordinates of the central point, width, height, and rotation angle) or eight parameters (coordinates of four vertices) to describe the rotated bounding box and l1 loss as the loss function. In this paper, we argue that the aforementioned integration can cause training instability and performance degeneration. The main reason is the discontinuity of loss which is caused by the contradiction between the definition of the rotated bounding box and the loss function. We refer to the above issues as rotation sensitivity error (RSE) and propose a modulated rotation loss to dismiss the discontinuity of loss. The modulated rotation loss can achieve consistent improvement on the five parameter methods and the eight parameter methods. Experimental results using one stage and two stages detectors demonstrate the effectiveness of our loss. The integrated network achieves competitive performances on several benchmarks including DOTA and UCAS AOD. The code is available at https://github.com/yangxue0827/RotationDetection.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/7322a4580299b322223730ca714a0a951788853a

How researchers see this paper

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

In Aerospace Engineering

1. Wentao Wang
2. Qi Ming
3. Junchi Yan
4. Qi Tian
5. Xue Yang
6. Yue Zhou
7. Liping Hou
8. Xiaopeng Zhang

In Computer Vision and Pattern Recognition

1. Qi Tian
2. Luc Van Gool
3. Junchi Yan
4. Ming-Hsuan Yang
5. Wangmeng Zuo
6. Lei Zhang
7. Dacheng Tao
8. Chen Change Loy

Citations in the corpus

• SCRDet++: Detecting Small, Cluttered and Rotated Objects via Instance-Level Feature Denoising and Rotation Loss Smoothing (2020, methodology): “Gliding Vertex [48] and RSDet [47] achieve more accurate object detection via quadrilateral regression prediction.”
• Localization Distillation for Object Detection (2022, methodology): “To address the boundary discontinuity and square-like problems, SCRDet [74] and RSDet [50] propose IoU-smooth L1 loss and modulated loss respectively for attaining smoother boundary loss, and CSL [72] proposes to use angle classification instead of angle regression.”
• Arbitrary-Oriented Object Detection with Circular Smooth Label (2020, methodology): “Although the parametric regression-based rotation detection method has achieved competitive performance in different vision tasks, and has been a building block for a number of excellent detection methods, these methods essentially suffer the discontinuous boundaries problem [30,44].”
• Rethinking Rotated Object Detection with Gaussian Wasserstein Distance Loss (2021, methodology): “For aerial images, ICN [2], ROI-Transformer [8], SCRDet [69] and Gliding Vertex [62] are two-stage representative methods whose pipeline comprises of object localization and classification, while DRN [41], R 3 Det [67] and RSDet [42] are single-stage methods.”
• Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence (2021, methodology): “Despite the theoretical grounds and the promising experimental justifications, our method has an obvious limitation that it cannot be directly applied to quadrilateral detection [33, 44].”
• MMRotate: A Rotated Object Detection Benchmark using PyTorch (2022, background): “…2 , 𝜋 / 2 ) ) [3, 6], and long edge 135° ) ) [5]. https://github.com/dingjiansw101/AerialDetection 4 https://github.com/jbwang1997/OBBDetection Recent works used two-dimensional Gaussian distribution [37, 39] and point sets [22–24, 31 ] to represent objects , which have achieved excellent results.”
• Poly Kernel Inception Network for Remote Sensing Detection (2024, methodology): “Some others [4, 27, 72, 73] utilize Gaussian distributions to model the OBBs for object detection and de-sign new loss functions [51] to guide the learning process.”
• Dense Label Encoding for Boundary Discontinuity Free Rotation Detection (2020, methodology): “Gliding Vertex [44] and RSDet [30] achieve more accurate object detection through quadrilateral regression prediction.”
• Large Selective Kernel Network for Remote Sensing Object Detection (2023, methodology): “In contrast with Oriented RCNN and Gliding Vertex, RSDet [54] addresses the discontinuity of regression loss by introducing a modulated loss.”