PixelScanner is our solution to automated 3D modeling from images. It is the end product of many years of research and is one of the most efficient solutions in the world. It is capable of processing hundreds of images to output Calibration information, a Dense 3D Point Cloud Model and a High Quality Textured Mesh Model of a scene in mere minutes without any user intervention.

It is performed in 4 distinct stages. The first stage is the capturing of the images. This is the only part where the user is involved in the process. The second stage, addressed by PixelScanner-Poser software, is the image calibration part where the geometric and photometric information about the data set is extracted. The third stage, addressed by PixelScanner-Cloud software, is the extraction of a dense point cloud model of the object. This stage makes use of the calibration information and requires heavy computational resources. The final and the fourth stage, performed by PixelScanner-Mesh, is the computation of a triangulated mesh from the point cloud and the texturization of this mesh.

See the Gallery for example models.

ex ex ex ex
ex ex ex ex

Example images for the Sinan Dataset


Calibration is the process of finding the true parameters of a camera that produces a certain given image. The camera parameters include the internal parameters like the focal length, principal point and lens distortion of the aquisition device and the external parameters that encode the geometric position and orientation of the device in the world.

Our PixelScanner-Poser software, by using image information alone, finds both the extrinsic and intrinsic parameters and outputs a sparse model of the scene. It is also possible to extend the calibration to metric in case it is necessary to make measurements on the final 3D model. Additionally, if GPS information is available, geo-referenced calibration is generated automatically.


A point cloud model is a set of vertices in 3D space representing an object. It is the most basic form for model representation and they are most often created by 3D laser scanners. These scanners operate by bouncing a laser beam on the surface of an object and output a point cloud as a data file.

Our PixelScanner-Cloud software emulates this process by only making use of the captured images and generates a very dense oriented point cloud model without using any active device. This means that users can generate a point cloud model of any textured object* by taking pictures of it with their cameras even from very far away distances like from a helicopter or from unmanned drones**. Using image data alone makes it possible to generate models of objects in a wide variety of settings for which the use of a laser scanner is not feasible or very expensive.

Our software not only generates colored dense point clouds but also extracts statistics like sampling density, curvature and surface normal information at each vertex location. This allows to use the generated models directly in graphics animations or industrial metrology applications or to use them directly to make digital elevation models. Moreover, it is even possible to generate models with a user-specified sampling density or with a preset number of points, or to downsample or upsample a given point cloud without losing the surface structure. Therefore, it is suitable to generate low resolution but high quality models for memory sensitive tasks.

*   See the Gallery for example models.
** See the Gallery for models estimated using images captured by UAVs and helicopters.


A triangle mesh model is a collection of vertices and triangle faces that defines an object in 3D computer graphics. It is the most commonly used model representation in games, animations and any application that makes use of 3D models. There is even customized hardware in today's computers, called GPU's, to process and render these triangulated mesh models.

Because of this popularity, we also provide triangulated mesh models of objects. Our PixelScanner - Mesh software automatically generates high quality textured mesh models of objects from their extracted point cloud models. Additionally, our software allows to control the resolution of the texture and the mesh independently and thus allows the user to reduce rendering memory and processing requirements in any way appropriate.