Point Cloud(s) / Photographic Modelling

Point clouds and photographic modelling are major topics that relate to modern use of instrument, digital and New Media perspective.

Point Clouds

A point cloud is a dataset representing objects or spaces as millions of discrete dots in 3-D space. Captured via LiDAR or photogrammetry, each dot contains precise X, Y, and Z coordinates and often color (RGB). Together, these points form highly accurate digital twins of the real world.

Photographic modelling, often known as photogrammetry, is the process of converting overlapping 2D images into accurate 3D models and point clouds. Every point in the cloud represents a precise location in 3D space, capturing real-world geometry, colour, and texture

The eye measures the apparent angular size of objects. A cloud represents a 3-D object with data points in space, each having coordinates and optional data like colour and timestamps. 

How are point clouds created? 

  • 3-D scanning: A 3-D scanner, like a laser scanner or LiDAR (Light Detection and Ranging), is
    used to gather data about a surface or structure.
  • Photogrammetry: Used to capture a building or site in colour, greyscale, or intensity.
  • Coordinate measuring machines: Data is collected about the outer profile of a 3-D object.
  • CAT scans: Data is collected about the inner structure of an object.

Applications

  • Architecture: Used to capture a building’s layout and current conditions.
  • Construction and restoration: Point clouds can be used to provide measurement data.
  • Manufacturing: Used to create 3-D models for product refinement and production.
  • Reverse engineering: Used to create surface models that can be converted into solids.
  • Robot navigation: Used to help robots navigate and perceive their environment.
  • Advanced driver assistance systems (ADAS): Point clouds can be used to help with depth estimation, stereo vision, and visual registration.

Digital Elevation Model

A Digital Elevation Model (DEM) is a 3-D representation of a planetary surface’s topography, showing “bare earth” by removing objects. DEMs are crucial for GIS analysis, hydrological modelling, and infrastructure planning, derived from sources like satellite radar, LiDAR, and topographic maps. 

Key Aspects 

  • Definition: DEMs represent bare earth, while DSMs include surface features.
  • Data Sources: Generated from NASA’s SRTM, ASTER, and LIDAR surveys.
  • Applications: Used in flood risk mapping, infrastructure, agriculture, and forestry.
  • Visualisation & Processing: Techniques like hill-shading and slope analysis enhance terrain.
  • Errors: Issues include “sinks” and “peaks,” needing correction for hydrologic modelling.

Common Types

  • SRTM: NASA radar data for 80% of land.
  • ASTER GDEM: Global DEM from satellite imagery.
  • Copernicus DEM: High-res global DEM at 30m resolution.
  • LiDAR DTM: Accurate ground elevation from laser scanning.

Orthophoto

An orthophoto is a geometrically corrected aerial or satellite image that removes distortions, merging photo detail with map precision for accurate measurements.

Key Aspects

  • Accuracy: Orthophotos have a constant scale for direct mapping.
  • Orthorectification: Corrects distortions using digital elevation models (DEM).
  • True Orthophoto vs. Traditional: True orthophotos fix building lean for vertical views;traditional ones do not.
  • Usage: Vital in GIS, surveying, construction, agriculture, and urban planning.
  • Resolution: High-resolution orthophotos from drones or aerial surveys (10-25 cm).

Applications

  • Engineering: Project planning and site mapping.
  • GIS: Creating base maps.
  • Agriculture: Monitoring crops.
  • Environment: Mapping coastal changes.

Gaussian Splat

Gaussian Splatting (3DGS) is a rasterisation-based, real-time rendering technique that creates photorealistic 3-D scenes from sparse 2-D images. It uses millions of tiny, transparent, 3-D ellipsoids (“splats”) containing colour and transparency data, enabling high-fidelity, high-speed rendering that outperforms Neural Radiance Fields (NeRF)s in speed and matches them in quality. 

Key Aspects of Gaussian Splatting 

  • Technology: Unlike mesh-based photogrammetry, Gaussian splatting uses millions of small,
    semi-transparent ellipsoids to represent 3-D spaces, offering high-fidelity and real-time aspect based colour dependent viewing.
  • Key Features: It uses an anisotropic Gaussian representation, enabling tailored lighting that captures complex reflections and light interactions.
  • Performance: It allows for interactive, real-time rendering, even in browsers.
  • Applications: Highly useful for digital asset showcases, virtual walkthroughs, and complex
    reality capture, such as on platforms like Polycam.
  • Workflow: Creating a scene involves capturing images from multiple angles, processing them, and then refining the Gaussian splat, a process similar to traditional 3-D modelling.

Limitations

  • While effective, Gaussian splatting may produce lower-quality results in less-seen or less- scanned areas, as with other rendering technologies.
  • The technology is new, with ongoing research in optimisation of core methods/systems.

Gaussian Splat – How it Works 

Gaussian Splatting (3DGS) : How It Works (Step-by-Step) 

1. Sparse Point Cloud Generation: The process starts by taking multiple photos of a scene and using a technique called Structure from Motion (SfM)—often similar to photogrammetry—to estimate camera positions and create a sparse 3-D point cloud. 

2. Gaussian Initialisation: Instead of relying on polygons, the system places 3-D Gaussians (ellipsoids) onto these sparse points. Each ellipsoid is defined by: 

  • Position ( x, y, z): Where it is in 3-D space.
  • Shape/Size ( covariance): How it stretches or rotates to cover surfaces.
  • Colour ( spherical harmonics): How colours change based on viewing angle • Opacity (α): How transparent or opaque it is. 

3. Iterative Optimisation: Through a training loop (similar to machine learning), the Gaussian parameters are adjusted to minimise the difference between the rendered scene and the original images. The density of the Gaussians is also managed: smaller Gaussians are added in fine detail areas, while unnecessary ones are removed. 

4. Real-time Rendering (Rasterisation): The optimised 3-D Gaussians are projected onto a 2-D image plane, sorted by depth, and blended together (alpha-blended) by a Graphics Processing Unit or GPU to produce a smooth, continuous, and highly detailed 3-D image in real-time.