| Title |
Automatic Image Segmentation Labeling Methodusing 3D Object Detection Dataset |
| Authors |
박석준(Seokjun Park) ; 박현욱(Hyunwook Park) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.10.89 |
| Keywords |
Image segmentation; Labeling; 3D object detection; Dataset |
| Abstract |
Recently, as stability and reliability have become important along with great advances in autonomous driving technology, research on 3D object detection is becoming more necessary. As a result, 3D object detection research is developing from single-modal object detection using only point cloud to multi-modal object detection using both image and point cloud. In multi-modal 3D object detection using image and point cloud, image segmentation, which includes object class information for each image pixel, can be useful to fuse two data. However, obtaining image segmentation labels requires a lot of time and cost, so the most open dataset of 3D object detection only have labels for bounding boxes, not image segmentation. Thus, this paper proposes how to automatically generate image segmentation labels using the bounding box label of the 3D object detection dataset. And by comparing image segmentation networks, one is trained using the generated image segmentation labels, the other is trained using human-made image segmentation labels, it was confirmed that the generated image segmentation label can replace the human-made image segmentation labels. |