The International Journal of Robotics Research Study, Volume 43, Issue 8, Web Page 1151-1174, July 2024.
Automation in farming is an expanding location of research study with basic social value as farmers are anticipated to generate even more and far better plant with less sources. An essential allowing element is robot vision strategies permitting us to feeling and afterwards connect with the atmosphere. A restricting element for these robot vision systems is their cross-domain efficiency, that is, their capability to run in a huge series of settings. In this paper, we recommend making use of supporting jobs to boost cross-domain efficiency without the demand for added information. We execute experiments making use of 4 datasets (2 in a glasshouse and 2 in cultivatable farmland) for 4 cross-domain examinations. These experiments show the performance of our supporting jobs to enhance network generalisability. In glasshouse experiments, our method enhances the panoptic top quality of points from 10.4 to 18.5 and in cultivatable farmland from 16.0 to 27.5; where a rating of 100 is the most effective. To better examine the generalisability of our method, we execute an ablation research making use of the huge Plant and Weed dataset (CAW) where we enhance cross-domain efficiency (panoptic top quality of points) from 12.8 to 30.6 for the CAW dataset to our unique WeedAI dataset, and 21.2 to 36.0 from CAW to the various other cultivatable farmland dataset. Although our recommended methods substantially enhance cross-domain efficiency we still do not usually outperform in-domain educated systems. This highlights the possible area for renovation around and the value of cross-domain research study for robot vision systems.
发布者：Michael Halstead，转转请注明出处：https://robotalks.cn/a-cross-domain-challenge-with-panoptic-segmentation-in-agriculture/
