Clarke, A.P., Haas, R., Hawemann, F., Toy, V.G 2026
Systematic Evaluation of the Influence of Camera Placement and Settings on the Quality of Photogrammetric Models of Rocks
Abstract
Despite the widespread adoption of photogrammetry across diverse disciplines, the relative influences of image acquisition parameters on the quality of photogrammetric models are seldom quantitatively understood. To address this, we conducted experiments under controlled lighting conditions, camera positions, and camera settings and evaluated the quality of the resultant models using both a subjective rating and a quantitative comparison. In total, 2541 models were evaluated in this study. In general, higher quality models can be produced by minimising large changes in the direction of view between adjacent images. Strong digital noise due to high ISO is detrimental to model quality, although this may be partially mitigated by noise reduction post-processing. RAW images generally produce higher quality models than JPEG images; however, at high ISOs, RAW images may result in poorer quality models due to their inherent lack of pre-applied noise reduction. Images taken with a smartphone produced models of comparable quality to those taken with a dedicated camera. Models were not consistently reproducible, even with near-identical images; therefore, practitioners must be aware of their margins of error when interpreting photogrammetric results. This study therefore provides practical guidance for practitioners based on a robust parameter study using natural geological samples.
Summary
Photogrammetry is a widely-used technique for 3D reconstructions of real-world objects and environments and is commonly used in geosciences to make virtual outcrop models for teaching and quantitative structural or morphological research. It is a relatively simple and cheap way to make a 3D model as it requires only a series of overlapping photos and a mostly-automated software workflow. Full-featured photogrammetric software includes Agisoft Metashape (commercial) and MeshRoom (open source).
The automated nature of this workflow has allowed many practitioners to achieve high-quality results without understanding the underlying processes of the photogrammetric software. As such, myths as to the optimum camera settings, camera arrangements, and software settings are perpetuated widely within practitioner communities. In this paper, we describe a systematic series of experiments with empirically validate the optimum settings and arrangements of cameras for photogrammetry and explain the underlying reasons why these parameters improve photogrammetric results. Through this paper, we hope to improve awareness and understanding of how photogrammetry works “under the hood” and enable practitioners to make decisions which improve the quality and reliability of their photogrammetric results.
Our research demonstrates that minimising the difference in perspective between adjacent camera positions — thereby requiring more photos to fully cover the object or scene — has the largest positive effect on the quality and reliability of photogrammetric results of any of the studied factors. We also found that minimising digital noise through the use of a low ISO also improved model quality. RAW images produced higher quality models at low ISOs than JPEGs, but at high ISOs the in-camera noise reduction applied to the JPEG images resulted in better models than the RAW images. Models made from images taken with a smartphone camera were only slightly better than those taken with a dedicated camera (Olympus E-M5 Mk. III). We also found that models produced from near-identical images were not of consistent quality, especially when a small number of widely-spaced cameras was used.
The issue of reproducability and precision of photogrammetric results is under-discussed in geoscience communities and should be taken into account, especially in studies which involve repeated surveys of the same area, such as ground-motion studies. These studies should quantify their margins of error in each survey before attributing differences in their models with changes in reality.
For this study, we took 46,944 images of three samples with two cameras and produced a total of 2541 models. These photos were taken using a stationary camera and a turntable to produce regularly-spaced stable camera positions. Lighting and exposure settings were controlled and the images were focus-stacked to remove blur from shallow depth of field. We then evaluated the quality of these models using a subjective rating system (giving each model a score from 0 to 6) and quantitate comparisons with a reference model.
Selected figures
Graphical Abstract
Graphical abstract showing influence of camera position and settings on photogrammetric model quality.
Model quality ratings for all models constructed
Model quality ratings for all models constructed.
Network graphs of matched image pairs
Network graphs showing matched image pairs between different camera positions in each of the camera network configurations.
Perspective distortion at different camera positions
Demonstration of the perspective distortion seen between images at different camera positions illustrating the distortion that would affect all feature points on the sample.
Comparisons between successful reconstructions
Selected photogrammetric models demonstrating successful reconstructions.