<img height="1" width="1" style="display:none;" alt="" src="https://px.ads.linkedin.com/collect/?pid=3382404&amp;fmt=gif">
Skip to content

Maximize measurement accuracy with images overlaid on point clouds

No quod sanctus instructior ius, et intellegam interesset duo. Vix cu nibh gubergren dissentias. His velit veniam habemus ne. No doctus neglegentur vituperatoribus est, qui ad ipsum oratio. Ei duo dicant facilisi, qui at harum democritum consetetur.
3d point cloud with measurement generated from photogrammetry software
The accuracy of a 3D measurement depends both on the accuracy of the reconstructed model and the ability of the user to select the right points on that model. The most common method to perform 3D measurements from photographs is to use photogrammetry to estimate 3D points and then fit a surface to those points to create an orthophoto or mesh. The problem is that the surface is simplified to fill holes and improve visualization. While the texture on the mesh gives the impression of detail, the geometry of the mesh is typically oversimplified and can be inaccurate.

Our approach is to use both the original images and the reconstructed points directly to perform measurement. Our recent user study shows that measurement using images overlaid on 3D point clouds with Reconstruct's viewer provides the best combination of precision and user experience, improving the accuracy by a factor of 4 vs. using point clouds alone and a factor of 2 vs. using a mesh.

Birds-eye view of scene used for benchmarking ground and vertical measurements. In this study, we evaluate the accuracy of linear measurements, such as the length of a parking space or a row of windows on a building.

Using the same set of photos (taken by drone), we created 3D models using Reconstruct's engine and, for comparison, Pix4D's engine. The photos were taken at a height of 100 ft and, for lateral shots, distance of 30 ft from the building. The resulting models are shown below, using screenshots from the Reconstruct and Pix4D viewers.

3D models created by the photogrammetry engines of Reconstruct (top) and Pix4D (bottom). The point clouds created by Reconstruct avoids the simplifying assumptions involved in mesh creation, leading to fewer inaccuracies and artifacts (right side). We have worked hard to make Reconstruct's 3D processing engine the most robust and accurate engine available for vertical structures like buildings and bridges.

We asked five users to perform several measurements on the ground and on the building walls using Reconstruct's 3D point cloud directly, images overlaid on the 3D point cloud, and Pix4D's online textured 3D mesh and measurement tools. We determined the true values by taking measurements directly in the scene.

Comparison of measurements performed without and with image overlay using Reconstruct's viewer. Measuring directly on the point cloud (left) can be difficult due to limited resolution and detail in the 3D model. Measuring on the images (right) provides the maximum possible resolution in perfect detail.

As expected, users report that overlaying the image makes it much easier to perform measurements. But we were surprised that using the image overlay improved accuracy compared to point cloud only by a factor of four, from an average error of 4" to 1" per measurement! The interface for 3D measurement is just as important as the precision of the 3D model for obtaining accurate measurements. This is an important observation because most reality capture services emphasize the theoretical limits of 3D point precision, rather than the precision of measurements obtained by users.

Results of our experiments. Five users each performed eight measurements using Reconstruct's web interface (with Reconstruct's 3D model) or Pix4D's interface (with Pix4D's 3D model). Some users tended to be more accurate than others, but performing measurements with images overlaid on the 3D model using Reconstruct's 3D web viewer resulted in twice the accuracy of measuring directly on 3D mesh models with Pix4D and four times the accuracy of measuring directly on point clouds (not shown). Errors shown are average error in inches or as a percent of total length.

In summary, the accuracy of your measurements from 3D models depends on both the interface and the precision of the model. Mesh models provide textured surfaces that are easy to interpret, but often oversimplify the surface or contain errors. Point clouds, by themselves, lack the high resolution of original images. Only Reconstruct provides image overlay in a web-based system to provide maximum detail and precision, making measurement simpler and more accurate.

Getting Started with Reconstruct

Want to learn how Reconstruct can help your project? Visit www.reconstructinc.com to start or with one of our team members.