ON March 22, a recent paper by Bodhidharma Institute was selected for CVPR 2020, the top computer vision conference. The paper proposes a general-purpose, high-performance automatic driving detector, which for the first time achieves both accuracy and speed of 3D object detection, effectively improving Autopilot system safety performance. Currently, the detector ranks first on the KITTI BEV leaderboard, an authoritative dataset in the field of autonomous driving.
The detector is the core component that enables autonomous driving to have perception capabilities. It can quickly process and analyze multi-dimensional information collected by sensors and lidars, allowing the vehicle to identify objects in the surrounding environment and accurately locate the object’s position in the three-dimensional space. Both accuracy and speed are important indicators of the safety of autonomous driving systems. However, the current mainstream single-stage detectors and two-stage detectors cannot take into account the two indicators.
Bodhidharma Institute put forward a new idea in the paper. It is understood that DAMO Academy uses an auxiliary network to convert the voxel features in the single-stage detector into point-level features in the model training process, supplemented by supervision signals, and the auxiliary network does not need to participate in the calculation during the model inference process, and finally realizes the A combination of speed and precision.
The test results show that the detector ranks first on the KITTI BEV ranking, the authoritative dataset in the field of autonomous driving, with far more accuracy than other single-stage detectors, and a detection speed of 25FPS, which is more than twice that of the current second-ranked solution. .
The team of the paper said, “The innovation of detectors is a key breakthrough in the field of autonomous driving. The detector we propose this time combines the advantages of a single-stage detector and a two-stage detector, so it achieves both 3D detection precision and speed. Improvement, innovative research on detectors in the future can also solve more problems in the autonomous driving industry.”
It is understood that the authors of the paper are all from Alibaba Damo Academy. The first author is Chenhang He, a research intern of Damo Academy. Other authors include Hua Xiansheng, senior researcher of Damo Academy and IEEE Fellow, senior researcher of Damo Academy and IEEE Fellow. Fellow Zhang Lei et al.