Illumination Aware Attention Networks for 3D Object Detection

Three-dimensional (3D) object detection is critical in a perception system of the Internet of Vehicles (IoVs). Although existing approaches fuse point cloud and image data, they often neglect the impact of environmental factors across sensing modalities. Using illumination variation in an IoV system, we propose a novel multimodal perception system, named as an illuminationaware attention network (IAANet), which dynamically adjusts feature weights. In particular, we first extract illuminationaware features from images and leverage an attention mechanism to recalibrate image feature weights before fusing them with point cloud features. We also design an ideal illumination-weight function and incorporate an auxiliary loss to enhance training effectiveness. The framework effectively mitigates interference from degraded image data in extreme illumination, thus improving overall detection accuracy. Comprehensive experiments on the KITTI 3D public dataset demonstrate that our method achieves competitive performance compared to other state-of-theart 3D detection models.