Application of Deep Learning to 3D Point Clouds<\/h1><\/span><\/div><\/div><\/div>a Supper & Supper Use Case<\/p>\n<\/div><\/div><\/div><\/div><\/div><\/div><\/div><\/div>
![](\"https:\/\/supperundsupper.com\/wp-content\/uploads\/2019\/10\/Deep-Learning-3D-Punktewolke-560x420.jpg\" "\\"Deep-Learning-3D-Pointcloud-560×420\\"")
<\/span><\/div><\/i><\/i><\/span>Project Goal<\/span><\/a><\/h4><\/div>\n3D point clouds are an enabling technology in the digitalization of many industries. Unlocking the information-richness of 3D point clouds requires additional analysis. This use case presents a scalable approach that applies neural networks to 3D point clouds to generate an instance segmentation of large scenic point clouds.<\/div>\n<\/div><\/div><\/div><\/i><\/i><\/span>Dataset used<\/span><\/a><\/h4><\/div>\nFor this use case we worked with photogrammetric point clouds generated by our drone flight partner FairFleet. The point clouds we analyzed were scans from urban and semi-urban environments, often containing construction sites. Each scan had up to ten million points.<\/div>\n<\/div><\/div><\/div><\/i><\/i><\/span>Challenges<\/span><\/a><\/h4><\/div>\nMost applications so far are limited to classifying single point clouds, for example 3D scans of individual objects. In order to properly make use of point clouds in for example construction and planning, it is necessary to segment large scenic scans with tens of millions of points.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Applied Methods (Implementation)<\/span><\/a><\/h4><\/div>\nTo segment large scenic 3D point clouds, we used a combination of unsupervised clustering and supervised neural network training.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Project outcome <\/span><\/a><\/h4><\/div>\nWe were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
![](\"https:\/\/supperundsupper.com\/wp-content\/uploads\/2019\/10\/Deep-learning_autoCAS_600x400.jpg\" "\\"Deep-learning_autoCAS_600x400\\"")
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>
a Supper & Supper Use Case<\/p>\n<\/div><\/div><\/div><\/div><\/div><\/div><\/div><\/div>
<\/span><\/div><\/i><\/i><\/span>Project Goal<\/span><\/a><\/h4><\/div>\n3D point clouds are an enabling technology in the digitalization of many industries. Unlocking the information-richness of 3D point clouds requires additional analysis. This use case presents a scalable approach that applies neural networks to 3D point clouds to generate an instance segmentation of large scenic point clouds.<\/div>\n<\/div><\/div><\/div><\/i><\/i><\/span>Dataset used<\/span><\/a><\/h4><\/div>\nFor this use case we worked with photogrammetric point clouds generated by our drone flight partner FairFleet. The point clouds we analyzed were scans from urban and semi-urban environments, often containing construction sites. Each scan had up to ten million points.<\/div>\n<\/div><\/div><\/div><\/i><\/i><\/span>Challenges<\/span><\/a><\/h4><\/div>\nMost applications so far are limited to classifying single point clouds, for example 3D scans of individual objects. In order to properly make use of point clouds in for example construction and planning, it is necessary to segment large scenic scans with tens of millions of points.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Applied Methods (Implementation)<\/span><\/a><\/h4><\/div>\nTo segment large scenic 3D point clouds, we used a combination of unsupervised clustering and supervised neural network training.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Project outcome <\/span><\/a><\/h4><\/div>\nWe were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>
<\/i><\/i><\/span>Dataset used<\/span><\/a><\/h4><\/div>\nFor this use case we worked with photogrammetric point clouds generated by our drone flight partner FairFleet. The point clouds we analyzed were scans from urban and semi-urban environments, often containing construction sites. Each scan had up to ten million points.<\/div>\n<\/div><\/div><\/div><\/i><\/i><\/span>Challenges<\/span><\/a><\/h4><\/div>\nMost applications so far are limited to classifying single point clouds, for example 3D scans of individual objects. In order to properly make use of point clouds in for example construction and planning, it is necessary to segment large scenic scans with tens of millions of points.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Applied Methods (Implementation)<\/span><\/a><\/h4><\/div>\nTo segment large scenic 3D point clouds, we used a combination of unsupervised clustering and supervised neural network training.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Project outcome <\/span><\/a><\/h4><\/div>\nWe were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>
<\/i><\/i><\/span>Challenges<\/span><\/a><\/h4><\/div>\nMost applications so far are limited to classifying single point clouds, for example 3D scans of individual objects. In order to properly make use of point clouds in for example construction and planning, it is necessary to segment large scenic scans with tens of millions of points.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Applied Methods (Implementation)<\/span><\/a><\/h4><\/div>\nTo segment large scenic 3D point clouds, we used a combination of unsupervised clustering and supervised neural network training.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Project outcome <\/span><\/a><\/h4><\/div>\nWe were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>
Most applications so far are limited to classifying single point clouds, for example 3D scans of individual objects. In order to properly make use of point clouds in for example construction and planning, it is necessary to segment large scenic scans with tens of millions of points.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Applied Methods (Implementation)<\/span><\/a><\/h4><\/div>\nTo segment large scenic 3D point clouds, we used a combination of unsupervised clustering and supervised neural network training.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Project outcome <\/span><\/a><\/h4><\/div>\nWe were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>
To segment large scenic 3D point clouds, we used a combination of unsupervised clustering and supervised neural network training.<\/p>\n<\/div><\/div><\/div>
<\/i><\/i><\/span>Project outcome <\/span><\/a><\/h4><\/div>\nWe were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>
We were able achieve accuracy scores of up to 93% for individual object classes, depending on the available labeled examples. The output of our network constitutes a so called \u201cinstance segmentation\u201d of the input point cloud. That means it is not only possible to assign a label to each 3D point (\u201csegmentation\u201d), one can also isolate individual instances of each label class. This additional step enables further analyses such as automatically counting and locating individual objects of a certain type.<\/p>\n
Our technology can turn point cloud data into valuable information, enabling you to successfully implement your digitalization strategy.<\/p>\n
Potential applications include urban infrastructure inventories, solar suitability analyses, construction site monitoring and 3D model generation<\/p>\n
Our 3D AI solutions can be adapted to your custom requirements and can be integrated into end-to-end solutions that feed results back into geodatabases, map layers, and AutoCAD models.<\/p>\n
<\/span><\/div>\n<\/div><\/div><\/div><\/div><\/div>