CVPR2023 | Phase-Shifting Coder: Predicting Accurate Orientation in Oriented Object Detection
paper:
https://arxiv.org/abs/2211.06368.pdf
code:
https://github.com/open-mmlab/mmrotate/tree/dev-1.x/configs/psc
Oriented object detection, also known as rotated object detection, is the task of detecting the position of an object while simultaneously obtaining information about its orientation. Different from horizontal object detection, oriented object detection usually uses rotated rectangles or arbitrary quadrilaterals for object detection, which can provide localization information more accurate and can be widely applied in the fields of scene text, remote sensing, industrial inspection, and so on. This article introduces a new method PSC for the boundary problem in oriented object detection, based on the CVPR2023 paper “Phase-Shifting Coder: Predicting Accurate Orientation in Oriented Object Detection”.
In this paper, a novel angle coder named phase-shifting coder (PSC) is proposed to accurately predict the orientation of objects, along with a dual-frequency version (PSCD). As we know, a vector can be represented as (x, y) coordinates or as length plus angle, and these two representations can be converted to each other. In oriented object detection, the angle of the object is quite similar to the vector. The PSC angle coder proposed in this paper treats the angle of the object as the vector angle and uses the periodicity of the trigonometric function to solve the boundary problem.
A rectangular box should be the same as a 180-degree rotated one, whereas a square box should be equivalent to under 90-degree rotation. Thus, by mapping the rotational periodicity of different cycles into the phase of different frequencies, common problems in oriented object detection such as boundary discontinuity and square-like problems are further solved in a unified form.
PSC is now supported in mmrotate 1.x. You are most welcome to evaluate it and report your results. And don’t forget to try PSC in designing a new oriented object detection method — there is a chance that you can get better performance.
