Improving Point Cloud Segmentation using Deep Learning

May 14, 2024

If you work in domains like mapping & surveying, infrastructure planning, or inspection, you must be familiar with the versatility of digital 3D modeling. Most of these applications require accurate segmentation of real-world objects to aid decision-making. Unfortunately, this process can be time-consuming and costly, as manual correction is often necessary to compensate for the low classification accuracy from “automated” tools.

To understand this better, consider the case of a city council that needs to survey a large area near a river for irrigation planning. Surveyors collect thousands of images using drones which are then processed using photogrammetry software to generate a 3D point cloud. From there, the point cloud must be segmented into different regions or objects like roads, buildings, water bodies, trees, vegetation cover, among others. This classification allows for better analysis and project management. 

Comparison of original and point cloud segmented image generated using Photogrammetry Platform
Graphical depiction of an area before and after point cloud segmentation. Different colors represent feature classifications like roads, buildings, ground, and structures.

But why is accurate point cloud segmentation so important?

Pointcloud segmentation is useful for extracting specific data points from a site. For instance, it can be used to calculate the number of buildings, extract the area covered by vegetation, and much more. It all boils down to the following functionalities:  

  • Accurate bare terrain representation – Point cloud segmentation can be used to accurately classify ground planes vs. non-ground features like buildings, trees, and more. This allows the generation of a unique representation with the geographical coordinates and the corresponding elevation, which is known as a Digital Terrain Model (DTM). This encoded information helps in analyzing bare terrain without the non-ground features. DTMs are used in civil engineering, geographic surveying, remote sensing, and urban planning. 
Digital Surface Model, non-ground features removed and Digital Terrain Model generated using Photogrammetry platform - Preimage
  • Understanding the data points of a site – Point cloud segmentation is also used to understand the area distribution and position of buildings, roads, vegetation/forest cover, and more. This information is majorly utilized in urban/rural planning and irrigation management.    
  • Autonomous driving and robotics – In autonomous cars, point cloud segmentation can be helpful in classifying various objects on roads and in the surrounding area, enabling the vehicle to determine the appropriate driving actions.  
  • Identifying areas of interest for high-resolution meshing – Point cloud segmentation can be used to identify regions of interest in a point cloud which can then be processed into a textured mesh at a higher resolution. For example, building or fine structures can be converted into a 3D textured mesh at higher resolution, compared to areas of lower interest with flatter geometry (e.g. ground or road).  

Unfortunately, the process of segmentation itself is not simple and hassle-free despite its numerous use cases in different industries. 3D segmentation is challenging due to noise, high redundancy, and uneven sampling density. Many existing tools that process point clouds are not entirely immune to the challenges and manual correction is often required to achieve the desired level of accuracy.

A GIS analyst who generally works on correcting the segmentation of point clouds to reach a certain accuracy can earn north of $60,000 per year. Especially for large survey sites, the process can be manually extensive and costly.  But what if this process could be made efficient with technology?  

Average salary of a GIS analyst in the USA
Source: Glassdoor

Deep Learning in Point Cloud Segmentation

These days, AI is increasingly being used to solve problems across business sectors, and it has been shown to outperform traditional methods for point cloud segmentation.

Large-scale mapping and surveying projects that cover extensive areas typically produce massive volumes of point cloud data, and traditional approaches may experience difficulties when it comes to promptly processing and analyzing this data. On the other hand, deep learning algorithms can process large amounts of data quickly and accurately, making them well-suited for point cloud segmentation in the mapping and surveying industry.

Another key advantage is the ability to handle complex or noisy data. Source data can be noisy or incomplete due to factors such as reflections, occlusions, or sensor noise. Deep learning algorithms can handle these issues better than traditional techniques, resulting in more accurate segmentation results.

Original Point Cloud, Ground truth segmentation & Point cloud segmentation with Deep Learning using Preimage
(a) Original Point Cloud, (b) Ground truth segmentation & (c) Point cloud segmentation with Deep Learning using Preimage

Using deep learning, it is possible to save hours of manual post-processing work, which can cost an average of $35/hour. Moreover, it allows analysts to focus their time on more value-added services like simulations and advanced analytics.  

Preimage’s 3D reconstruction platform uses deep learning techniques for point cloud segmentation and delivers up to 90% classification accuracy. You can try out our AI-enabled 3D segmentation web application (no installation needed). When you register your account, you will receive 10,000 credits preloaded, which is adequate for segmenting 50 square kilometers of RGB point cloud.

You might like this!

See More

Try Preimage’s 3D Models

View our reconstruction engine in action with visualizations of virtual tours, textured meshes and 3D point clouds created with Preimage AI.