摘要
由于室内植物的叶片存在大量自遮挡,为了得到植物的完整三维信息,往往需要用户手动裁剪、扫描和配准叶片.针对该问题,提出了利用实例分割网络进行叶片识别并选取被裁剪叶片的方法.通过总结植物叶片形状与分布特征,建立虚拟植物模型,渲染大量带叶片轮廓信息的图片来训练实例分割网络,避免了耗费大量精力标注真实植物叶片.进一步提出了自动化建模系统,根据单个视角下观察到的植物图像自动选择被裁剪的叶片,并获得该叶片的三维模型;调整相机位置让更多的叶片被检测到并剪掉,以恢复植物的三维信息.使用该系统对绿萝叶片、鸭脚木和花烛3种虚拟植物进行了测试,并评估重建的整株植物重合度、叶片重合度和叶片占比,结果显示使用本文方法重建的植物模型能更完整地恢复原始植物的三维信息.
Due to the heavy self-occlusion of indoor plants,to capture the complete 3D information of a plant,users usually need to manually cut the leaves,scan them,and register the leaf scans together.In response to this problem,we propose a method to identify and select the leaves to crop using the instance detection network.To avoid manually labeling photos of real plant leaves for training,we summarize the leaf shape and distribution to build 3D plant models,then render lots of images with leaf contour information.An automatic intrusive reconstruction system for capturing the full 3D of indoor plants is introduced,which automatically and successively selects a leaf to cut from a captured image to obtain its 3D points and adjusts the camera view when nothing is detected.Three types of virtual indoor plants have been tested and evaluated.Results show that the reconstructed plant models capture the plants well.
作者
史川石
郑倩
黄惠
Shi Chuanshi;Zheng Qian;Huang Hui(Visual Computing Research Center,Shenzhen University,Shenzhen 518060)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2021年第2期161-168,共8页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(61861130365,U2001206)
广东省自然科学基金(2020A0505100064,2015A030312015)
广东省高等学校科技创新重点项目(2018KZDXM058)
深圳市基础研究基金(JCYJ20180305125709986).
关键词
植物建模
侵入式建模
叶片检测
实例分割
plant modeling
intrusive modeling
leaf detection
instance segmentation
