Research on an efficient and deployment-friendly 3D object detection algorithm

The deployment of a 3D detector is one of the main challenges in autonomous driving applications. From the perspective of industrial applications, this paper proposes an efficient and deployment-friendly 3D object detection algorithm, OptimPillars, to improve the accuracy of small object detection and optimize the deployment process. First, we introduce average attention pooling and propose the Max-and-Avg Attention Pillar Encoding (MAAPE) encoding method, which is specifically used to improve the detection effect of small objects. Secondly, by redesigning the backbone network and optimizing the allocation of computing resources, the goals of high performance and low latency are achieved. Finally, in the neck module, by introducing the ASPP module, multi-scale dilated convolution is used to expand the receptive field, which effectively improves the multi-scale fusion effect of features and enhances the model's perception of contextual information. Experimental results show that OptimPillars has demonstrated excellent performance on both the nuScenes dataset and the self-built dataset.