ARIoU: Anchor-free Rotation-decoupling IoU-based optimization for 3D object detection

Abstract

The rotation-decoupling strategy was developed in outdoor 3D object detection with certain performance improvement. However, its anchor-based architecture limits its further improvement in indoor 3D object detection. In this study, we propose an Anchor-free Rotation-decoupling IoU-based optimization, termed ARIoU, which is specifically devised to indoor 3D object detection. A pivotal aspect of our approach is a customized encoding-decoding scheme that integrates the anchor-free framework with the rotation decoupling strategy to mitigate the negative effect caused by rotation sensitivity. Specifically, the unwanted length/width/rotation prediction randomness induced by labeled rotation ambiguity is alleviated. Furthermore, we plug our ARIoU as the optimization targets in both classification and regression branches to solve the misalignment issue in predicting precision and confidence estimation. Experiments on the SUN RGB-D and S3DIS datasets demonstrate the effectiveness of our approach. Code is available at https://github.com/wenchened/ARIOU.