Object Detection in Cluttered Environments With Sparse Keypoint Selection

Viktor Seib, Dietrich Paulus; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 2496-2505

In cases such as mobile robotic applications with limited computational ressources, traditional approaches might be preferred over neural networks. However, open source solutions using traditional computer vision are harder to find than neural network implementations. In this work we address the task of object detection in cluttered environments in point clouds from RGB-D cameras. We compare several open source implementation available in the Point Cloud Library and present a novel and superior solution for this task. We further propose a novel sparse keypoint selection approach that combines the advantages of uniform sampling and a dedicated keypoint detection algorithm. Our extensive evaluation shows the validity of our approach, which also improves the results of the compared methods. All code is available on our project repository: https://github.com/vseib/point-cloud-donkey.