Rethinking Boundary Discontinuity Problem for Oriented Object Detection

TL;DR

A dual-optimization paradigm for angles is proposed, which for the first time achieves the objectives of both correcting angular boundary and blending angle with other parameters, and extensive experiments show that boundary discontinuity problem is well-addressed.

Abstract

Oriented object detection has been developed rapidly in the past few years, where rotation equivariance is crucial for detectors to predict rotated boxes. It is expected that the prediction can maintain the corresponding rotation when objects rotate, but severe mutation in angular prediction is sometimes observed when objects rotate near the boundary angle, which is well-known boundary discontinuity problem. The problem has been long believed to be caused by the sharp loss increase at the angular boundary, and widely used joint-optim IoU-like methods deal with this problem by loss-smoothing. However, we experimentally find that even state-of-the-art IoU-like methods actually fail to solve the problem. On further analysis, we find that the key to solution lies in encoding mode of the smoothing function rather than in joint or independent optimization. In existing IoU-like methods, the model essentially attempts to fit the angular relationship between box and object, where the break point at angular boundary makes the predictions highly unstable.To deal with this issue, we propose a dual-optimization paradigm for angles. We decouple reversibility and joint-optim from single smoothing function into two distinct entities, which for the first time achieves the objectives of both correcting angular boundary and blending angle with other parameters.Extensive experiments on multiple datasets show that boundary discontinuity problem is well-addressed. Moreover, typical IoU-like methods are improved to the same level without obvious performance gap. The code is available at https://github.com/hangxu-cv/cvpr24acm.

Figures (7)

• Figure 1: Two optimization paradigms for angle in oriented object detection: (a) in joint-optim methods yang2021rethinkingyang2021learningyang2023kfiouzhou2019iou, smoothing function is explicitly applied for detector's output $\theta_p$ during loss calculation; (b) while in independent-optim methods yang2020arbitraryyu2023psc, smoothing function is implicitly embedded in the model, and $\theta_p$ is decoded from detector's output $f_p$. According to our analysis, only the latter can really solve boundary discontinuity problem.
• Figure 2: When objects rotate near the boundary angle, state-of-the-art IoU-like methods (e.g., KFIoU yang2023kfiou, KLD yang2021learning) actually suffer from severe mutation in angular prediction. With the correction for angle by our ACM, the prediction achieves rotation equivariance.
• Figure 3: Box $\neq$ Object: (a) objects rotated with 45$^\circ$ and 225$^\circ$[colorful mark] share the same box rotated with 45$^\circ$ [black mark], which causes (b) the relationship [blue line] between $angle$ of box and object to become a piecewise function with a breakpoint [gray region], differing from the $(position, scale)$. Not only is the prediction [red line] of the breakpoint region mutational, but the prediction of other regions also becomes fluctuant.
• Figure 4: Overview of proposed Dual-Optimization paradigm and ACM-Coder. The detector outputs angular ACM-encoding $f_p$, subject to explicit supervision. On this basis, another IoU-like loss based on joint-encoded $g(\cdot)$ is applied onto ACM-decoded angle $f^{-1}(f_p)$. The paradigm achieves the objectives of both correcting angular boundary and blending parameters.
• Figure 5: Visualized comparison of detection results between KFIoU yang2023kfiou and enhanced ACM-KFIoU. The images are arranged from left to right in order of increasing aspect-ratio of objects, and the first-col and bottom-col are the results of KFIoU and ACM-KFIoU, respectively. Our ACM greatly eliminates the angular prediction errors in the original KFIoU.