对植物幼苗进行三维重建,常存在叶片间的遮挡、摄像头视野限制等因素导致植物幼苗点云出现缺失的情况,影响了植物表型分析的准确度。为了能获得完整的植物点云,提出了基于多尺度几何感知Transformer(Multi-Scale Geometry-Aware Point T...对植物幼苗进行三维重建,常存在叶片间的遮挡、摄像头视野限制等因素导致植物幼苗点云出现缺失的情况,影响了植物表型分析的准确度。为了能获得完整的植物点云,提出了基于多尺度几何感知Transformer(Multi-Scale Geometry-Aware Point Transformer,MGA-PT)的植物点云补全网络。该网络首先通过降采样特征提取模块对原始点云进行邻域特征提取;然后利用Transformer提取语义信息,引入多尺度几何感知模块提取不同尺度下的几何信息,加强对植株不同器官的特征提取能力;最后使用双路稠密点云生成模块分别对输入部分和预测部分进行细粒度生成,避免输入点云特征的丢失,保证稠密点云贴近实际分布。试验使用基于运动恢复结构的方法对植物幼苗进行三维重建,通过旋转与固定视点缺失构建数据集。试验结果表明,该补全网络表现出色,比目前主流的补全网络更优,对植株数据集补全结果的倒角距离为0.79×10^(-4)cm,地面移动距离为0.11 cm,F1分数为70.77%,且对不同形态、不同比例的缺失均能补全,体现网络具有稳定性与健壮性。该网络对叶类植物补全效果好,为植物幼苗点云补全提供了新思路。展开更多
For many tller crops,the plant archit ecture(PA),including the plant fresh weight,plant height,number of tllrs,tller angle and stem diameter,sigificantly afects the grain yield.In this study,we propose a method based ...For many tller crops,the plant archit ecture(PA),including the plant fresh weight,plant height,number of tllrs,tller angle and stem diameter,sigificantly afects the grain yield.In this study,we propose a method based on volumetric reconstruction for high-throughput three-dimensional(3D)wheat PA studies.The proposed methodology involves plant volumetric reconst ruction from multiple images,plant model processing and phenotypic parameter estimation and analysis.This study was performed on 80 Triticum aestium plants,and the results were analyzed.Comparing the automated measurements with manual measurements,the mean absolute per-centage error(MAPE)in the plant height and the plant fresh weight was 2.71%(1.08cm with an average plant height of 40.07cm)and 10.06%(1.41g with an average plant fresh weight of 14.06 g),respectively.The root mean square error(RMSE)was 137 cm and 1.79g for the plant height and plant fresh weight,respectively.The correlation cofficients were 0.95 and 0.96 for the plant height and plant fresh weight,respectively.Additionally,the proposed methodology,in-cluding plant reconstruction,model processing and trait ext raction,required only approximately 20s on average per plant using parallel computing on a graphics processing unit(GPU),dem-onstrating that the methodology would be valuable for a high-throughput phenotyping platform.展开更多
文摘对植物幼苗进行三维重建,常存在叶片间的遮挡、摄像头视野限制等因素导致植物幼苗点云出现缺失的情况,影响了植物表型分析的准确度。为了能获得完整的植物点云,提出了基于多尺度几何感知Transformer(Multi-Scale Geometry-Aware Point Transformer,MGA-PT)的植物点云补全网络。该网络首先通过降采样特征提取模块对原始点云进行邻域特征提取;然后利用Transformer提取语义信息,引入多尺度几何感知模块提取不同尺度下的几何信息,加强对植株不同器官的特征提取能力;最后使用双路稠密点云生成模块分别对输入部分和预测部分进行细粒度生成,避免输入点云特征的丢失,保证稠密点云贴近实际分布。试验使用基于运动恢复结构的方法对植物幼苗进行三维重建,通过旋转与固定视点缺失构建数据集。试验结果表明,该补全网络表现出色,比目前主流的补全网络更优,对植株数据集补全结果的倒角距离为0.79×10^(-4)cm,地面移动距离为0.11 cm,F1分数为70.77%,且对不同形态、不同比例的缺失均能补全,体现网络具有稳定性与健壮性。该网络对叶类植物补全效果好,为植物幼苗点云补全提供了新思路。
基金supported by grants from the National Program on High Technology Development(2013AA102403)the Program for New Century Excellent Talents in University(NCET-10-0386)+1 种基金the National Natural Science Foundation of China(30921091,31200274)the Fundamental Research Funds for the Central Universities(2013PY034).
文摘For many tller crops,the plant archit ecture(PA),including the plant fresh weight,plant height,number of tllrs,tller angle and stem diameter,sigificantly afects the grain yield.In this study,we propose a method based on volumetric reconstruction for high-throughput three-dimensional(3D)wheat PA studies.The proposed methodology involves plant volumetric reconst ruction from multiple images,plant model processing and phenotypic parameter estimation and analysis.This study was performed on 80 Triticum aestium plants,and the results were analyzed.Comparing the automated measurements with manual measurements,the mean absolute per-centage error(MAPE)in the plant height and the plant fresh weight was 2.71%(1.08cm with an average plant height of 40.07cm)and 10.06%(1.41g with an average plant fresh weight of 14.06 g),respectively.The root mean square error(RMSE)was 137 cm and 1.79g for the plant height and plant fresh weight,respectively.The correlation cofficients were 0.95 and 0.96 for the plant height and plant fresh weight,respectively.Additionally,the proposed methodology,in-cluding plant reconstruction,model processing and trait ext raction,required only approximately 20s on average per plant using parallel computing on a graphics processing unit(GPU),dem-onstrating that the methodology would be valuable for a high-throughput phenotyping platform.
