Close-Range Photogrammetry and 3D Recording

SfM is often described as a push-button method because the software does most of the mathematical work automatically. That description is misleading. The software can only compute a reliable model if the image set it receives meets specific conditions. Images that are too sparse, that lack overlap, that have a uniform texture the software cannot match, or that are sharp in some views and blurred in others will produce a degraded or unusable model. The skill is in planning and executing the image capture, not in clicking Metashape buttons.

1. Plan the coverage
   Define the area to be modelled and decide on a camera grid or walking path that ensures 80 to 90 percent overlap in both directions. For a grave 2 m by 1 m, a grid of overlapping vertical shots from 0.5 to 1.5 m height covers the floor; angled shots around the perimeter capture the section walls.
2. Place and measure GCPs
   Before photography begins, distribute at least four GCPs across the area (six to eight is better for larger scenes or where high accuracy is required). Photograph the GCPs at close range to confirm they are identifiable in the images. Measure their coordinates with the total station or RTK-GPS.
3. Capture images
   Shoot in RAW format where possible; use a fixed focal length lens; avoid wide apertures that blur the depth of field; keep ISO low to minimise noise. In low-light conditions (deep trenches, overcast days) use a tripod or pole-mounted camera and adjust shutter speed rather than raising ISO above 800.
4. Process in Metashape
   Import images, align cameras (SfM step), assign GCP coordinates to their image locations, build a dense point cloud, generate a textured mesh and orthophoto, export at the required resolution and coordinate system.
5. Quality check
   Review the reprojection error report. Typical acceptable values are below 0.5 pixels for the GCPs and below 1 pixel overall. Compare a spot measurement in the model against the same measurement from the total-station record. Differences over 10 mm indicate a processing or GCP error that needs investigation.

Feature matching, the core of SfM alignment, works by finding the same physical point in two or more different images taken from different viewpoints. For that to work, the point must appear in enough images (which depends on overlap), must appear at a recognisably different angle in at least two of those images (which requires camera positions that are not all on the same vertical line), and must be identifiable rather than part of a featureless uniform surface.

• Overlap requirement: 80 percent along the direction of travel and 60 percent sideways. In practice this means adjacent images share about two-thirds to four-fifths of their frame content.
• Nadir plus oblique combination: a first pass of vertical (nadir) shots models the horizontal surface. A second pass of oblique shots at 45 degrees captures section walls and the edges of the grave cut where vertical shots produce insufficient parallax.
• Diffuse lighting: hard shadows from direct sunlight create surface features in one image that disappear in another, confusing the matcher. Overcast conditions or a reflector panel to fill shadows gives more reliable matching. If bright sun is unavoidable, shoot consistently with the sun at the same angle relative to the camera.
• Featureless surfaces: clean sand or wet clay can be nearly textureless. Applying chalk dust or placing small numbered tokens across the surface introduces artificial features the software can match. This is standard practice for recording exposed bone surfaces.

SfM can reconstruct shape from photographs alone, but the scale, orientation, and absolute position of the model are arbitrary unless external control is supplied. GCPs are the connection between the photogrammetric model and the total-station coordinate system that holds the rest of the site record.

Physical GCPs are typically printed coded targets (Metashape can auto-detect its own code patterns) or hand-made crosses on A4 card, weighted flat and placed to avoid movement during photography. They should be distributed around the margins of the modelled area and across its interior, not clustered in one corner. Four is a minimum; six to eight gives redundancy for checking.

In Metashape, GCP assignment is done in the Reference panel. Each marker is placed manually on its image position in every photograph where it appears, or auto-detected if coded targets are used. Metashape then optimises the camera positions and model geometry with the GCP coordinates as constraints. The final report shows the residual at each GCP in x, y, and z and the root-mean-square error (RMSE) across all GCPs. For forensic work an RMSE below 5 mm is typically expected.

Metashape produces several complementary outputs from the same image set, and each has a different role in documenting and presenting the evidence.

For court presentation the orthophoto is usually the most accessible product. It can be printed at 1:10 or 1:20 with a scale bar, annotated with context numbers, and presented alongside the hand-drawn context plans as a photographic check on their accuracy. The mesh and point cloud are more often submitted as supplementary digital exhibits or used by opposing experts for independent measurement.

Terrestrial laser scanning (TLS) captures a scene by firing a laser in a dense regular grid from a tripod-mounted scanner and measuring distance by time-of-flight or phase-shift. A single scan from one position takes 5 to 20 minutes and produces tens of millions of points. Multiple scans from different positions are registered to a common frame using targets or cloud-matching algorithms.

• Accuracy: TLS achieves 2 to 6 mm per point over distances up to 50 m. SfM with good GCPs achieves 3 to 10 mm per derived point. For most forensic grave recording the difference is not operationally significant.
• Speed in the field: TLS is faster at capturing a scene but requires each scan position to be set up carefully. SfM image capture is fast (a complete grave can be photographed in 10 to 20 minutes) but post-processing is computationally intensive and takes longer.
• Equipment cost: a mid-range TLS unit costs 30,000 to 100,000 USD; a camera capable of forensic SfM work costs 500 to 3,000 USD. This makes SfM the dominant choice for routine forensic archaeology, with TLS reserved for major scenes.
• Low-light performance: TLS is light-independent and works in darkness. SfM requires adequate illumination for photograph quality. For underground scenes or night operations, TLS is preferred.

Courts in several jurisdictions have accepted SfM-derived 3D models and orthophotos as exhibits, but the admissibility argument rests on a few specific foundations. The practitioner must be able to state: how many images were taken and what overlap was used; how many GCPs were placed and what their measurement accuracies were; what software version was used and what its published accuracy claims are; what the RMSE of GCP residuals was; and whether an independent check point confirmed the stated accuracy. These are not optional additions: they are the chain of custody for the spatial data.

A limitation that courts sometimes probe is the risk of model artefacts: areas of the point cloud or mesh where the reconstruction is unreliable because of image gaps, featureless surfaces, or moving objects during photography. These show up as holes, floating point clusters, or unrealistically smooth surfaces in the model. Processing reports should identify such areas, and measurements should not be taken from them. Masking unreliable regions in the orthophoto before submission is standard practice.