For safe driving of autonomous vehicles, it is crucial to detect 3D objects within point clouds with both real-time performance and accuracy. Currently, many autonomous driving research groups adopt simple and efficient pillar-based 3D detectors for real-time performance. However, these detectors often apply grids of large size, which can potentially lead to the loss of significant point information. Moreover, these detectors commonly rely on post-processing steps such as NMS and pseudo-label generation. This reliance can adversely impact detection accuracy because it does not fully reflect the object’s location information. To address these issues, we propose 3DSimDet, an efficient and compact 3D object detection framework suitable for autonomous driving platforms. This framework comprises two key components. The proposed SimBackbone module is designed to enhance the feature encoding capabilities of traditional methods. BIoU Head module includes a classification branch and an IoU prediction branch, which considers the object’s location information during inference. Moreover, we introduce a high-quality pseudo-label generator for semi-supervised learning, leveraging the prediction outcomes from the BIoU Head module during the semi-supervised learning phase. Through extensive experimentation on diverse public autonomous driving datasets such as WOD, H3D, and A3D, we demonstrate the effectiveness of our proposed method. The experimental findings demonstrate that our 3DSimDet exhibits superior overall performance in terms of accuracy and runtime compared to conventional pillar-based supervised learning models. Furthermore, the pseudo-label generator proves to enhance detection performance within a semi-supervised learning framework.

LiDAR-based 3D object detectors usually adopt grid- based approaches to handle sparse point clouds efficiently. However, during this process, the down-sampled features inevitably lose spatial information, which can hinder the detectors from accurately predicting the location and size of objects. To address this issue, previous researches proposed sophisticatedly designed neck and head modules to effectively compensate for information loss. Inspired by the core insights of previous studies, we propose a novel voxel-based 3D object detector, named as Re-VoxelDet, which combines three distinct components to achieve both good detection capability and real-time performance. First, in order to learn features from diverse perspectives without additional computational costs during inference, we introduce Multiview Voxel Backbone (MVBackbone). Second, to effectively compensate for abundant spatial and strong semantic information, we design Hierarchical Voxel-guided Auxiliary Neck (HVANeck), which attentively integrates hierarchically generated voxel-wise features with RPN blocks. Third, we present Rotation-based Group Head (RGHead), a simple yet effective head module that is designed with two groups according to the heading direction and aspect ratio of the objects. Through extensive experiments on the Argoverse2, Waymo Open Dataset and nuScenes, we demonstrate the effectiveness of our approach. Our results significantly outperform existing state-of-the-art methods. We plan to release our model and code1in the near future.

Recently, due to the increasing interest in autonomous driving, many researchers have been focused on Open Source Software (OSS) such as the Autoware platform, which enables full-stack autonomous driving. As a research group dedicated to the development of autonomous driving software platforms, we are particularly focused on the development of 3D object detection models that take into account real-time performance as well as accuracy. In this paper, we propose a VariFocal-based CenterPoint model and an adaptive shape estimation method combined with Apollo segmentation to improve the performance of the 3D object detection module included in the Autoware platform. Additionally, we present a Bird’s Eye View (BEV) metric-based quantitative evaluation method to compare the performance of different models according to the dataset. By means of the evaluation method, we demonstrate the experimental results of the performance between the 3D object detection modules of the Autoware platform and our models on both our own dataset and a public dataset. Furthermore, since these models will be used for autonomous driving platforms where real-time processing is a crucial factor, we also perform experimental results on the processing time by converting the models to TensorRT. As shown in the experimental results, the improved models trained on a large amount of self-constructed datasets resolve issues such as missed detection of large vehicles and small objects like motorcycles and pedestrians compared to the previous models. As a result, VF-CenterPoint performed 13.76 mAP better than the previous model, and A-Shape Estimation method achieved 9.7 AP higher performance than the previous method.