Super-fast and accurate monocular 3D object detection for Autonomous Driving

Professor Dongsuk Kum and his team have developed a super-fast and accurate 3D object detection method using a single monocular camera. Their method achieved 1^st place on the KITTI monocular 3D object detection leaderboard (most widely adopted perception benchmark for autonomous driving) while running real-time, outperforming prior methods.

Article | Fall 2021

Detecting 3D objects in a 3D driving environment is an essential but very challenging perception task for autonomous driving. In particular, when only a single camera is used, detecting 3D object becomes very challenging because an image does not contain distance information. Therefore, retrieving lost 3D information from images plays a key role in monocular image-based 3D object detection.

 

Professor Dongsuk Kum and his team have developed a super-fast and accurate 3D object detection method using a single monocular camera without the need for a large amount of 3D bounding box annotations. The method to solve the data shortage problem proposed by the research team exploits raw LiDAR points as training data in addition to a 3D bounding box to train the depth information. This allows to avoid labelling instance-level 3D bounding boxes, which is a very expensive labelling task. The research team was able to dramatically improve computational time by using image-based detector coupled with an auxiliary depth prediction network, which focuses on predicting the distance of the detected objects.

 

The proposed method achieved 1^st place on the KITTI monocular 3D object detection leaderboard (most widely adopted perception benchmark for autonomous driving), outperforming the state-of-the-art methods by 12% while running three times faster (42ms). Moreover, it is advantageous that their method is not limited to a specific 3D detection network architecture but can be generally applied to many other networks. This research was published in IEEE Transaction on Intelligent Transportation System (T-ITS) under the title of “Boosting Monocular 3D Object Detection with Object-Centric Auxiliary Depth Supervision.