P²CBAM: 3D Object Detection for Autonomous Driving

Efficient and accurate object detection from 3D point clouds is crucial for autonomous driving. Due to the sparsity and irregularity of real-world point clouds, as well as the real-time requirements for detectors, existing 3D object detectors face significant challenges in practical applications. PointPillars is widely used in real-world scene due to its fast detection speed, but its accuracy lags behind some existing methods. To enhance the detection accuracy of this model, we propose P<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>CBAM, which integrates CBAM into the PointPillars baseline model. This integration improves detection accuracy with relatively minimal additional computational resources. Extensive experiments conducted on the KITTI and Waymo Open datasets demonstrate that P<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>CBAM outperforms the baseline model on both datasets. The code is available at https://github.com/Bennyliiiiii/P2CBAM.