PP-SSD: Point Painting Single Stage Detector for 3D Object Detection
Yulin Liu, Huihui Helen Wang, Hao Liu, Shiyou Chen · 2024
Integrating multimodal data features as input for downstream modules would provide superior detection performance. Using an independent semantic model to paint point clouds constraints the performance of object detection. 2D image data only records the color information of foreground objects, using color information to enhance the features of all point clouds remains a challenging task. In our work, we proposed a simple and plug-and-play point painting method to tackle above issue. The color information of the point cloud in the same frustum is encoded using the range information before the point-based single-stage detector, integrating the position, intensity and color information. Experiments results on the KITTI datasets using 3DSSD network as point encoder validate the effectiveness of our point painting method. On the KITTI car 3D detection test, our PP-3DSSD achieves 82.71% AP.