FeatFlow: Learning geometric features for 3D motion estimation
作者：
Highlights：
• We propose a novel end-to-end learning based framework that predicts the dense scene flow and the ego-motion of the platform by establishing the effective geometric correspondence from consecutive point clouds.
• We introduce a new point-score layer and a score-weighted feature alignment loss function. Their combination effectively drives the model to learn distinctive keypoint detectors and consistent feature descriptors.
• On popular scene flow datasets (FlyingThings3D and KITTI Scene Flow), our framework surpasses all the state-of-the-art methods in scene flow estimation by a considerable margin. On popular real LiDAR datasets (Oxford RobotCar and KITTI Odometry), our framework also produces pairwise scan-matching results comparable to the state-of-the-art.
摘要
•We propose a novel end-to-end learning based framework that predicts the dense scene flow and the ego-motion of the platform by establishing the effective geometric correspondence from consecutive point clouds.•We introduce a new point-score layer and a score-weighted feature alignment loss function. Their combination effectively drives the model to learn distinctive keypoint detectors and consistent feature descriptors.•On popular scene flow datasets (FlyingThings3D and KITTI Scene Flow), our framework surpasses all the state-of-the-art methods in scene flow estimation by a considerable margin. On popular real LiDAR datasets (Oxford RobotCar and KITTI Odometry), our framework also produces pairwise scan-matching results comparable to the state-of-the-art.
论文关键词：Feature learning,Motion estimation,Point clouds,Scene flow,Scan-matching,Ego-motion
论文评审过程：Received 4 March 2020, Revised 17 June 2020, Accepted 4 August 2020, Available online 20 August 2020, Version of Record 29 September 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107574
