Recommandé
Recommandé
Contenu connexe
Similaire à pydataPointCloud.pptx
Similaire à pydataPointCloud.pptx (20)
Plus de Manuel Rodrigo Cabello Malagón
Plus de Manuel Rodrigo Cabello Malagón (11)
Dernier
Dernier (20)
pydataPointCloud.pptx
- 1. Enriching 3D point cloud data with Artificial Intelligence @mrcabellom mrcabellom@gmail.com Research Engineer @ Plain Concepts Rodrigo Cabello
- 2. The procedures to convert point clouds into application-specific deliverables are very costly in time/manual intervention. It is necessary to develop processes that extract the essential information automatically to create valuable data for decision-making.
- 3. 3D representation • Better understanding of our environment. 3D data can provide more dimensional information. • 3D models can: • Represent the features of virtually any object. • Represent complex objects with a finite number of elements. (Point Cloud) • Improve the decision-making process. • Design error reduction: • industry and building sector.
- 4. Data representation Point Cloud Voxels (Volumetric Pixels) Meshes Pixels 2D image 3D Model
- 5. RGB-D provides a 2,5D information 3D data representation
- 6. 3D data representation - NERF
- 7. Point cloud
- 8. • A point cloud is a set of data points in space. • Points may represent a 3D shape or object. • Each point position has its set of Cartesian coordinates (X, Y, Z). • Points can include attributes like RGB, intensity, and classification. • Point clouds are generally produced by 3D scanners (LIDAR) or by photogrammetry software. An example of a 1.2 billion data point cloud render of Beit Ghazaleh, What is a point cloud?
- 10. Millions of points Unstructured data Multiple file formats Noisy Sensors Challenges
- 11. Point cloud + AI • Automated insights extraction in large point clouds. • ML-assisted capacity can help reduce human errors by automatically pre-labeling. Classification, custom object detection, and segmentation. Object detection in the 2D image and 3D projection. Synthetic Data Generation Deep Learning
- 12. Object detection – Cube RCNN
- 13. Engineer Data Ingestion Data Cleansing/ Transformation Model Training/ Building Model Testing Model Deployment Test Data 1 2 3 4 4 4 5 Model feedback Loop Train Test Loop AI Workflow Version 1 Azure Machine Learning Model registry
- 14. • Point cloud • Data transformation to 3D voxel grid projections. • Data normalization (scale) • Translation, rotation, and permutation invariance. • Sort input into canonical order. • Large point clouds. Memory consumption. • Slices, segmentation • Downsampling Data preprocesing - Deep Learning
- 15. Data cleaning • Pass-through filter. • Statistical outlier removal. • Radius outlier.
- 16. Downsampling • Voxel Grid Filter. Centroid • Farthest point sample
- 17. Semantic segmentation Object detection and localization Segmentation Classification Computer vision tasks PointNet++ PointCNN SE-PseudoGrid PointPillars PointRCNN FCAF3D RANSAC Correspondence Grouping K-Means, DBSCAN PAConv PointNetTransformers
- 19. 01 Create the digital version of the environment. Automatic digital twin
- 20. 01 Create the digital version of the environment. Automatic digital twin 02 Extract main environment insights.
- 21. 01 Create the digital version of the environment. Automatic digital twin 02 Extract main environment insights. Detect and recognize specific objects. 03
- 22. 01 Create the digital version of the environment. Transform environment and insights into 3D objects. Automatic digital twin 02 Extract main environment insights. Detect and recognize specific objects. 03 04
- 23. 01 Create the digital version of the environment. Transform environment and insights into 3D objects. Automatic digital twin 02 Extract main environment insights. Detect and recognize specific objects. 03 04 IoT devices simulator 05
Notes de l'éditeur
- Extracting information from point clouds is very expensive. It requires a lot of time and manual intervention to get valuable insights into the industry sector (For example: generating BIM models…) The use of Artificial intelligence for object identification and 3D model creation will save time and will allow the creation of valuable data for decision-making.
- As you know All artificial intelligence models need data to learn and make predictions in the future, if we are using a system to classify images, we can represent an image with a 2D matrix… but what about 3D environments? We can represent 3d data as meshes (triangles and vertex.), voxels (volumetric pixels.), and point clouds (a set of data points in the space).
- Other techniques give us information about 3D space like images with depth
- like the Neural radiance field. This technique uses Deep Learning to create a 3D model from a set of 2D images.
- What is exactly a point cloud?
- How can we combine point clouds and AI?
- But.. What are the main challenges? Multiple file formats: e57, LAS, LAZ… Unstructured data… unordered points (X, y, Z) with RGB information. Point clouds are large, sometimes we can have a file with a billion points. The environment and sensor precision are essential to obtain quality data.
- As mentioned before, we can use AI to extract some information from point clouds. Deep Learning models can help us to extract lines, edges, or basic geometric forms to understand the environment and detect complex objects.
- Statistical outlier removal: Performing a statistical analysis on each point’s neighborhood, and trimming those which do not meet a certain criterion. Our sparse outlier removal is based on the computation of the distribution of point to neighbors distances in the input dataset. Radius outlier removal: Is a simple filter that removes outliers if the number of neighbors in a certain search radius is smaller than a given K.
- Statistical outlier removal: Performing a statistical analysis on each point’s neighborhood, and trimming those which do not meet a certain criterion. Our sparse outlier removal is based on the computation of the distribution of point to neighbors distances in the input dataset. Radius outlier removal: Is a simple filter that removes outliers if the number of neighbors in a certain search radius is smaller than a given K.
- PAConv -> Position Adaptative Convolution FCAF3D -> Fully convolutional Anchor-Free 3D Object Detection
- Ok, let’s see an example. From the research team, we have been working on a process to classify piping components from lidar point clouds using Deep Learning techniques. The use of AI is indispensable for model retrieval from point clouds, especially when creating 3D models of irregular objects for the scan-to-BIM process.