May 13, 2024 | Duke I3T Lab's paper appearing at IEEE ICRA 2024
While a large and growing body of work has been assessing the performance of 3D object detections on conventional LIDAR-based and RGB-D based point clouds, the performance of 3D ODs on VI-SLAM point clouds, commonly employed in augmented reality applications, has not received as much attention. VI-SLAM point clouds are, in many ways, unlike LIDAR or RGB-D point clouds: the accuracy, density, and spatial distribution of VI-SLAM data vary widely depending on object and background textures. In this work we create and release two VI-SLAM point cloud datasets that capture such environmental variability. Using these datasets, we quantify the differences between traditional and VI-SLAM point clouds and evaluate the performance of three leading 3D ODs, examining the extent to which object and environment characteristics affect their performance.
[Duan243D] L. Duan, Y. Chen, T. Scargill, and M. Gorlatova, 3D Object Detection with VI-SLAM Point Clouds: The Impact of Object and Environment Characteristics on Model Performance, in Proc. IEEE ICRA, Yokohama, Japan, May 2024. Paper PDF
Great thanks to the National Science Foundation and Meta for supporting this research!