Passive scanning
Table of contents
Photogrammetry
Photogrammetry is the collection and organization of reliable information about physical objects and the environment through the process of recording, measuring and interpreting photographic images and patterns of electromagnetic radiant imagery and other phenomena.
Photogrammetry was first documented by the Prussian architect Albrecht Meydenbauer in 1867. Since then it has been used for everything from simple measurement or color sampling to recording complex 3D Motion Fields
Unlike other scanning methods that require precise orbital plans or specialized equipment, photogrammetry can be achieved simply by flying a drone in a circular pattern and capturing multiple photos. Utilizing the location data from the drone, one can construct detailed models like the example shown here: A typical medium resolution aerial photogrammetry scan of a barn. With 50-100 images a reasonably accurate model can be produced. Such models are often used in surveying and restoration projects from the scale of hand helf objects to cities. This accessibility makes photogrammetry an attractive option for various applications, with results that can be sufficiently accurate depending on the specific requirements.
However, it’s essential to note that photogrammetry lacks inherent scale. Without a reference point or prior knowledge of the camera locations, the resulting model lacks a definitive scale, as cameras inherently lack absolute scale information. Therefore, incorporating at least one reference point is crucial. For example, marking a facade with visual markers or known distances, such as pieces of tape, allows for scaling within a 3D modeling program based on these references.
Stereo Matching
Stereo matching is also known as "disparity estimation", referring to the process of identifying which pixels in multiscopic views correspond to the same 3D point in a scene.
Early uses in stereophotogrammetry, the estimation of 3d coordinates from measurements taken from two or more images through the identification of common points. This technology was used throughout the early 20th century for generating topographic maps.
Cameras were applied to photogrammetry as early as 1973, such as in Imaging characteristics of photogrammetric camera systems , an early paper describing groundwork camera properties and standards for USGS (United States Geological Survey) photgrammetry surveys.
When describing photogrammetric camera systems, intrinsics and extrinsics parameters are used to characterize systems:
Intrinsic parameters describe the properties of a camera’s lens and sensor.
Extrinsic parameters describe the camera’s position and orientation in 3D space
Today: Levenberg–Marquardt algorithm or damped least-squares algorithm (dls) are used to minimize the error across 3d coordinates. This procedure is typically called bundle adjustment.
While the analog versions of these techniques have waned in popularity, stereophotogrammetry still has applications for capturing dynamic characteristics of previously difficult to measure systems like running wind turbines.
When is it useful? Photogrammetry is useful for outdoors settings, where all you need is a handheld camera and some patience. In this example, note the loss of quality towards the top, as pixel resolution becomes problematic:
Key Benefits
- Unified Workflow Geometry and texture/color in one workflow.
- Affordability and flexibility. Depending on the end use application almost any camera will work given there is enough light and your post processing software is robust.
- Real Time Feedback & Processing* *as models improve. Nasa’s Ingenuity Drone relies on photogrammetry-based onboard processing for ground distance estimation, showcasing the efficacy of passive sensing approaches in complex environments.
Key Challenges
-
Lighting Light conditions in the scene are crucial to the quality of the scan. A controlled environment is highly preferred. Precision is improving but can still be completely thrown off by certain light conditions in much the same way LiDar struggles with smooth surfaces.
-
Precision Limitations Increasingly industry pairs vision systems for photogrammetry with laser systems to balance the benefits of both.
Photogrammetry scan using Polycam, 2m range, 3 circular passes around the mannequin. ~2 minute process to get the result. On the left, rendered view, on the right, the untextured mesh.
Software
Open Source
AliceVision Open-source photogrammetric computer-vision framework
Meshroom is the 3D Reconstruction Software built on AliceVision
3D Software Suites
Autodesk ReCap
Agisoft Metashape
Pix4D - Uses ML for improving accuracy while scraping information on the contents of photogrammetry data sets.
Apps
iPhone and Android apps for photogrammetry and now LiDAR scanning have multiplied over the last several years:
Polycam
Rekrei
Adobe Aero
ScandyPro - Real-time meshing
As the apps multiply, scanned objects become more widely available for both art preservation and computer graphics applications:
Scan the World
Quixel Megascans
Light Field
plenoptic: Of or relating to all the light, travelling in every direction, in a given space.
Light fields represent an advanced form of passive sensing, aiming to capture full plenoptic content: all possible light rays emanating from a scene in any given direction. This results in a four-dimensional function, as it involves selecting a ray’s position and angle. If the ideal plenoptic function was known, any novel viewpoint could be synthesized by placing a virtual camera in this space, and selecting the relevant light rays.
“We begin by asking what can potentially be seen” - Edward H. Adelson & James R. Bergen, Media Lab, Vision & Modeling Group: The Plenoptic Function and the Elements of Early Vision, 1991.
Why do we want all of the light?
Image-Based Rendering (IBR) for view synthesis is a long-standing problem in the field of computer vision and graphics. Applications in robot navigation, film, and AR/VR.
The affect of this technique was popularized in The Matrix, in the famous bullet dodge shot which used 120 still cameras and two film cameras.
This is such a resource intensive setup, that it prompts researchers to seek simulation shortcuts to reach this result, such as this paper using thousands of virtual cameras and neural networks to capture a complete dense plenoptic function.
In practice, we can only sample light rays in discrete locations. There are two popular optical architectures for this:
Multi-Camera Systems
Simply shoot the scene from several locations using an array of camera (or a single moving one).
Some sophisticated acquisition rigs shown by Google in in a SIGGRAPH 2018 paper:
Lenslets
Lenslets: a single CMOS sensor with an array of lenses in front.
In the lenslet approach, each pixel behind a lenslet provides a unique light ray direction. The collection for all lenses is called a sub aperture image, and roughly corresponds to what a shifted camera would capture. The resolution of these images is simply the total number of lenslets, and the number of sub-aperture images available is given by the number of pixels behind a lenslet. For reference, the Lytro Illum provides 15x15 sub-aperture images of 541x434 pixels each, which is a total of ~53 Megapixels.
The most efficient layout for lenslets is hexagonal packing, as it wastes the fewest pixel area. Note that some pixels are not fully covered by the lenslet and receive erroneous or darker data. This means some sub aperture images cannot be recovered.
Depth Estimation
Forming an image from these cameras requires sampling one pixel from each micro lens to generate virtual viewpoints. The resulting “sub-aperture images” offer different perspectives with subtle shifts, presenting a challenge for depth estimation due to their minute disparities.
Depth estimation on Light Field data is an active domain. For now, algorithms are commonly tested on ideal, synthetic light fields such as this dataset. Here is one example of point cloud obtained from a stereo matching method.