Interpreting Photographs and Images

Some things are easier to see on a photograph than on a 3D model — a fracture network, a facies boundary, anything that reads as a pattern in the image. VRGS opens photographs and orthoimages in their own 2D Image view with a full interpretation toolset, and because the photographs are calibrated, what you draw there can be projected back onto the mesh.

Open a photograph by double-clicking it in the Data tree. The 2D Image ribbon tab appears.

Which tool for which job

You want	Use
To outline an object — a bed, a block, a clast	Segment
Every fracture trace in the image	Fracture
A trace between two points, following the rock	Image Trace or Segmentation Trace
To draw a trace yourself	Draw Line
To colour regions by facies	Paint
To pull out linear features from texture alone	ACO

Segment — Segment Anything

Segment runs the Segment Anything model over the image. Five ways to tell it what you want:

Mode	How you drive it
Points	Click inside the thing you want. Add more clicks to refine it, including negative clicks to push the boundary back out of somewhere it wrongly included.
Rectangle	Drag a box around the object.
Box (SAM 3)	The same, run through SAM 3.
Prompt	Type what you are looking for in words.
Automatic Segmentation	Segment everything in the image, producing a complete set of masks with no prompting at all.

Settings puts the segmentation parameters into the properties panel.

Choosing the model

SAM 2 model on the ribbon picks the SAM 2.1 Hiera variant:

Variant	Size	When
tiny	~150 MB	The default. Fast, and adequate for well-defined objects.
small	~600 MB	A step up at modest cost.
base_plus	~2 GB	Noticeably better on ambiguous boundaries.
large	~3.5 GB	Best quality.

The choice is saved per project, and a variant downloads lazily the first time it is needed — so switching to large mid-session costs a download before the next encode, not immediately.

Use SAM 3 for Points routes ordinary point clicks through SAM 3 when it is loaded for the current image, falling back to SAM 2 silently when it is not. It is off by default.

SAM 3 needs a CUDA GPU

The SAM 3 paths — Prompt, Box (SAM 3) and SAM-3-for-points — require a CUDA-capable GPU. Without one VRGS says so and falls back to SAM 2 click segmentation. See AI & Machine Learning Requirements.

Segmenting in 3D as well

Entering Segment mode also arms point picking in the 3D view, so you can click the feature on the model and have the points reproject into the photograph. That is what makes a label consistent across several overlapping photographs rather than being re-guessed in each one.

Fracture detection

Fracture runs the fracture-detection model over the image and extracts fracture traces. Fracture (Hi-Res) runs it at higher resolution for fine or faint traces, at proportionally more cost.

The traces land in the same editable 2D line set the manual tools use, so everything in the next section applies to them.

2D Line Settings puts the extracted-line parameters into the properties panel.

For the full workflow, tuning recipes and diagnostics, see AI Fracture Mapping.

ACO — Ant Colony Optimisation

ACO is the classical alternative to the fracture model: it finds linear features from image texture with no trained model involved. Useful where the AI model has not seen anything like your rock, and as a cross-check. Settings exposes its parameters.

See the ACO guide.

Editing the extracted lines

Everything the automatic tools produce — and everything you draw by hand — forms one editable line set. These tools work on all of it:

Tool	What it does
Select Line	Click to select; Shift-click to add. Delete erases the selection.
Erase Line	Click a line to delete it.
Erase Vertex	Remove a vertex; the line is dropped below two vertices.
Join Selected	Chain the selected lines into one by their nearest endpoints.
Split Line	Click an interior vertex to cut the line in two.
Move Vertex	Drag a vertex.
Insert Vertex	Click a segment to add a vertex at the nearest point on it.
Extend Line	Click an endpoint, then digitise onward; double-click to finish.
Draw Line	Place vertices for a new trace; double-click to finish.

The two auto-trace tools

Both take a start click and an end click and route the trace between them for you. The difference is what they follow:

Tool	Follows	Needs
Segmentation Trace	The fracture-probability field	Fracture to have been run first
Image Trace	The image's own edge and colour gradient	Nothing — works on any image

Use Image Trace for a bedding contact or any boundary the eye can see; Segmentation Trace where the AI has already found the fracture and you just want it cleanly connected.

Both are tuned by Trace Settings on the Interpretation panel.

Auto-join and hysteresis

Right-click in the image view for two toggles worth knowing: Auto-join Fragments on Extract joins broken fragments during extraction, and Hysteresis Thresholding grows traces from strong seeds so a sub-threshold dip along a fracture does not split it in two. Both reduce the amount of joining you have to do by hand.

Painting facies labels

Rather than outlining regions, you can paint them:

Control	What it does
Paint	Paint the active facies label onto the image with a brush.
Erase	Erase label pixels back to background.
Brush Size	Brush diameter in image pixels — a preset or a typed value.
Square Brush	Square footprint instead of round.
Label	The active facies, with a button to create a new one.

This is the quickest route to a training set, and to a coarse facies map where precise boundaries are not the point.

Other interpretation tools

Wand selects connected pixels of similar colour or intensity from one click. Polyline and Polygon digitise annotations on the image. Erase Point and Move Point edit their vertices. Tape Measure measures distances on the image.

Image processing

The Image Processing panel applies filters to help you see what you are interpreting — they change the display, not the stored image:

• Edge Detection, plus Vertical, Horizontal and Gabor variants for oriented features
• Brightness / contrast
• Contours — intensity contour lines over the image
• Filter Width, Iterations and Blend control the kernel and how far the result is mixed with the original
• Reset Image discards everything and restores the original

Graph plots pixel values along a profile, and Variogram computes one for the image or a selection.

Filter to see, not to interpret

An edge-detected view makes faint traces obvious. Trace them, then Reset Image — the interpretation is stored against the photograph, not against the filtered view.

Getting the result onto the 3D model

Because photographs are calibrated, a segmentation can be projected through the camera onto the mesh.

Right-click in the image view and choose Bake Labels to Mesh Texture. The labels are projected onto every mesh visible in that photograph and stored as an AI Classification attribute layer, which you can recolour and query like any other attribute. Repeating the bake after segmenting more photographs accumulates into the same layer rather than replacing it.

The entry is greyed out until the active photograph actually has a segmentation layer.

See Texture Attributes for what happens to the layer once it is on the mesh.

Display and navigation

Reset View, Zoom In, Zoom Out and Target Zoom on the Home panel; Flip and Rotate on the Display panel; and First / Previous / Next / Last on Navigate Images to step through the set.

Tips and troubleshooting

• Segmentation is slow the first time. The image has to be encoded before any prompt can be answered. Later prompts on the same image are fast.
• A prompt or SAM 3 box does nothing. Those paths need CUDA. VRGS reports the fallback rather than failing silently.
• The mask keeps including the background. Add a negative point in the part you want excluded, rather than starting over.
• Fracture traces come out fragmented. Turn on Hysteresis Thresholding and Auto-join Fragments on Extract, then re-run Extract. Join Selected cleans up what is left.
• Bake Labels to Mesh Texture is greyed out. The photograph has no segmentation layer yet — segment something first.
• Baked labels land in the wrong place on the mesh. The photograph's camera calibration is the projection; an unregistered photo cannot texture anything. Check the photograph is registered.

See also

• AI Fracture Mapping — the fracture workflow in full.
• ACO for Image Analysis — the classical detector.
• Texture Attributes — what happens to baked labels.
• AI & Machine Learning Requirements — GPU and model requirements.