A Compare Research of Two Different Point Clouds 3D Object Detection Methods

Object detection in point clouds serves as an important foundation for many applications such as autonomous driving and roadside perception.The existing methods for this foundation can be roughly divided into two categories, which are one-stage methods and multi-stage methods.For the one-stage method, an improved Pointpillars neural network, called MSCS-Pointpillars, was proposed to detect objects directly from point clouds.Here, an attention mechanism and pillars of different scales for the Pointpillars network were introduced to solve the problem of information loss caused by single scale pillar partition.For the multi-stage method, a flexible multi-stage algorithm AF3D, where point clouds were first clustered into clusters which were then detected by a much simpler classifier based on deep learning, was proposed.The two methods on both KITTI dataset and our own dataset have been compared.The results show that MSCS-Pointpillars exhibits better accuracy, but it is difficult to maintain its good performance in unfamiliar scenes.For AF3D, the accuracy appears worse, but it demonstrates much better robustness to unfamiliar scenes.