Learning Modulated Loss for Rotated Object Detection

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