平行点云: 虚实互动的点云生成与三维模型进化方法

三维信息的提取在自动驾驶等智能交通场景中正发挥着越来越重要的作用, 为了解决以激光雷达为主的深度传感器在数据采集方面面临的成本高、样本覆盖不全面等问题, 本文提出了平行点云的框架. 利用人工定义场景获取虚拟点云数据, 通过计算实验训练三维模型, 借助平行执行对模型性能进行测试, 并将结果反馈至数据生成和模型训练过程. 通过不断地迭代, 使三维模型得到充分评估并不断进化. 在平行点云的框架下, 我们以三维目标检测为例, 通过闭环迭代, 构建了虚实结合的点云数据集, 在无需人工标注的情况下, 可达到标注数据训练模型精度的72%.

Parallel Point Clouds: Point Clouds Generation and 3D Model Evolution via Virtual-real Interaction

The extraction of 3D information is playing an increasingly important role in intelligent traffic scenes such as autonomous driving. In order to solve the problems faced by LiDAR sensor such as the high cost and incomplete coverage of possible scenarios, this paper proposes parallel point clouds and its framework. For parallel point clouds, virtual point clouds are obtained by building artiflcial scenes. Then 3D models are trained through computational experiments and tested by parallel execution. The evaluation results are fed back to the data generation and the training process of 3D models. Through continuous iteration, 3D models can be fully evaluated and updated. Under the framework of Parallel Point Clouds, we take the 3D object detection as an example and build a point clouds dataset in a closed-loop manner. Without human annotation, it can be used to effectively train the detection model which can achieve the 72% of the performance of model trained with annotated data.