← Back to Omniris Omniris · 3DGS Field Guide

Capturing Drone Imagery for High-Quality Gaussian Splats

A practical field guide to flying for 3D Gaussian Splatting (3DGS) reconstruction. This tutorial assumes you’ve flown photogrammetry missions before — if you have, most of what you know transfers, but the priorities shift in important ways.

Why 3DGS isn’t just photogrammetry with a different output

Both pipelines start the same way: you fly a drone and collect overlapping photos. From there, the workflows diverge:

Three practical consequences flow from this:

  1. Lighting must be consistent across the entire capture. A splat trained on inconsistent lighting will produce “floaters” — wispy cloud-like Gaussians that try to reconcile irreconcilable views. Photogrammetry tolerates lighting drift; 3DGS does not.
  2. You need more views per surface, from more angles, than for photogrammetry. Because the model represents view-dependent appearance, every surface should be observed from a hemisphere of directions, not just from above and once obliquely.
  3. Underconstrained regions become artifacts, not just holes. Where photogrammetry produces a hole or low-poly patch, 3DGS hallucinates Gaussians that look correct from training views but break down when you move the camera off the captured trajectory.

Plan for these from the moment you walk onto the site.

Step 1: Pre-flight planning

Pick the right time of day

Diffuse, overcast light is the gold standard. It eliminates harsh shadows and the directional speculars that cause inconsistencies between orbit angles. If you must shoot in direct sun:

Survey the scene for failure modes

Walk or pre-fly the site and identify:

Choose your scene scale

The capture pattern depends on what you’re reconstructing:

These patterns are described below.

Step 2: Camera settings

Manual everything. Auto-exposure between frames is the single most common cause of bad splats. Even small EV shifts produce a flicker effect that 3DGS cannot resolve.

The settings that matter

Setting Recommendation Why
Exposure mode Full manual Auto-exposure variance ruins photometric consistency
Shutter speed 1/1000s minimum, 1/2000s preferred Eliminates motion blur. Drones move and vibrate; even 1/500s can blur fine detail at altitude
ISO Lowest you can while maintaining shutter Noise floor matters for SH coefficient quality
Aperture Fixed (most drone cameras have fixed aperture; honor it) If variable, lock at lens sweet spot (typically f/4–f/5.6)
White balance Manual / locked Auto WB drift causes color inconsistency between angles
Format RAW (DNG) if available, otherwise highest-quality JPEG RAW lets you correct exposure post-hoc and apply consistent processing across the set
Picture profile D-Log or flat profile if RAW unavailable Preserves dynamic range for downstream processing
Mechanical shutter Use it if your platform supports it Eliminates rolling-shutter distortion that confuses pose recovery

Shutter speed math

To keep motion blur below one pixel, the rough rule is:

shutter_speed < (GSD) / (drone_speed)

At 5 m/s with a 2 cm GSD, that’s 0.02 / 5 = 0.004s → 1/250s minimum just for forward motion. Add gimbal vibration headroom and you’re at 1/1000s. Fly slower if you can’t push shutter higher.

Pre-process before handoff

If you’ve shot RAW, develop the entire set with identical settings before handing the data off — same exposure, same WB, same tone curve. Don’t auto-balance per-image. The whole point is photometric consistency.

Step 3: Flight patterns

Pattern A — Building or single structure
Open visualizer →

The workhorse pattern. Combine these passes:

  1. Nadir grid over the structure and surrounding ground at altitude H. Side and forward overlap ≥ 80%.
  2. Orbit ring at H₁ ≈ 1.5× structure height, gimbal pitched -45°. One photo every 5–10° of azimuth (so 36–72 photos per ring).
  3. Orbit ring at H₂ ≈ structure height, gimbal pitched -20° to -30°. Same azimuth density.
  4. Low orbit at H₃ ≈ ⅓ structure height, gimbal near horizontal (-5° to -15°), capturing the base, façades, and any overhangs.
  5. Detail orbits for ornamentation, antennae, recessed features — closer-in passes at varied gimbal angles.

Modern platforms with automated oblique mission modes will run variants of this for you. Manual flight is fine; just hit the rings systematically.

Pattern B — Open site or landscape
Open visualizer →
  1. Double-grid nadir at altitude H — fly a grid in one direction, then orthogonal. 80% overlap both axes.
  2. Oblique grid at altitude H, gimbal pitched -45°, in the same flight direction as one of the nadir grids. This is the single most important pass for splat quality on flat sites.
  3. Perimeter orbit(s) around the area of interest at the edges of the site, capturing the scene from outside-in.

If your scene has vertical features (buildings, trees, terrain breaks), add Pattern A around each.

Pattern C — Tall structure (tower, monument, cliff face)
Open visualizer →

Skip the grid. Fly:

  1. Concentric perimeter orbits at 4–6 altitudes, evenly spaced from base to slightly above the top.
  2. Vary radius between rings — alternate close and far so you capture both detail and context.
  3. Vary gimbal pitch per ring — top rings tilt down more, bottom rings tilt up more, middle rings closer to horizontal. The combined coverage gives you the hemisphere of viewing directions every part of the structure needs.
  4. One nadir pass over the top to capture the crown, plus one wide nadir over the base for ground context.
Pattern D — Single object (small subject)
Open visualizer →

For a sculpture, vehicle, or individual tree:

  1. Treat it as a sphere. Capture from a dome of viewpoints — three orbits at three pitch angles (-15°, -45°, -75°) is a reasonable minimum.
  2. Distance: close enough that the object fills 60–80% of the frame.
  3. Supplement with ground-level handheld or tripod shots if the bottom and undersides matter.
Pattern E — Wind turbine (or other rotating radial structure)
Open visualizer →

Wind turbines combine tall-structure capture with radial blade geometry. They are also the pattern with the most consequential safety constraints in this guide.

The blades MUST be locked stationary for the entire capture. Even slight rotation between frames poisons multi-view consistency catastrophically — far worse than wind-blown vegetation, because the moving objects are large, repetitive, and dominate the scene. Confirm rotor lock visually and with the operator before takeoff.

The “Y” position (one blade pointing straight up) is conventional and gives symmetric coverage. With the rotor locked in Y:

  1. Hub-level orbit at extended radius — the critical pass. Altitude = hub height. Radius ≥ 1.5× blade length to maintain safety clearance from the lower blade tips. Gimbal near horizontal (-5° to -10°), capturing the hub and the inner two-thirds of all three blades from the side.
  2. High orbit at altitude = hub height + 20%. Smaller radius (≈ 1× blade length, since you’re now above the upper blade tip). Gimbal -30° to -45°, capturing the top of the hub and the upper blade from above.
  3. Mid-tower orbit at altitude ≈ ½ hub height. Radius ≈ 1× blade length. Gimbal -10°, for the tower midsection.
  4. Low orbit at altitude ≈ ¼ hub height. Radius small (~0.5× blade length, well clear of the lower blade tips, which sweep at altitude = hub_height − blade_length). Gimbal +15° looking up, for the tower base and lower hub.
  5. Nadir pass directly over the hub at a comfortable vertical margin. Gimbal -90°. Captures the hub crown, sensors, and any wiring on top of the nacelle.
  6. Ground supplementation for foundation, connection box, transformer pad, and any access infrastructure.

Critical safety notes for wind turbines:

Step 4: Coverage rules of thumb

These are the heuristics worth internalizing. Use them as starting points and adjust based on results from your specific scenes:

Step 5: Ground supplementation

Aerial-only captures fail at the bottoms of structures and anywhere a roof, awning, or overhang occludes the view from above. For high-fidelity results — especially on buildings with façade detail — supplement with terrestrial capture:

This single addition often produces a larger quality jump than any flight-pattern tweak.

Step 6: Capture-side preparation

A few preparation tasks that pay off in final splat quality, all done before handing imagery to the Omniris 3D Splatting Pipeline:

Step 7: Quality assessment

After processing, evaluate the splat in a viewer and look specifically for:

If the splat fails any of these checks badly, the fix is usually in the capture, not the training. Re-fly with the gap in mind.

Quick reference: the field checklist

Pre-flight
  □ Diffuse light or stable sun position
  □ No precipitation, low wind
  □ Manual exposure locked
  □ Manual white balance locked
  □ Shutter ≥ 1/1000s
  □ RAW or highest-quality JPEG
  □ Mechanical shutter enabled if available
  □ Battery count: aerial pattern + ground supplement + 20% margin

Mission
  □ Nadir grid (80%/80% overlap)
  □ Oblique grid OR multi-altitude orbits (per scene type)
  □ Low/horizontal pass for façades and undersides
  □ Detail orbits for high-frequency features
  □ Ground supplementation walk

Post-flight
  □ Cull blurry frames
  □ Cull redundant poses
  □ Develop RAW with consistent settings
  □ Verify metadata integrity
  □ Hand off to Omniris 3D Splatting Pipeline

The single best habit you can build is treating the first capture of any new site as a calibration flight. Fly it, run a quick splat that night, and look at where it fails. Then fly the real mission knowing what your scene specifically needs more of. Capture quality compounds — a small improvement in coverage, lighting consistency, or sharpness pays off many times over in the final result.

References

Foundational research

Drone capture for 3DGS specifically

Field practice references

Note on heuristics

Several specific numbers in this guide — image count targets, the orbit altitude formulas (1.5× height, ⅓ height), the three-altitude minimum, and the camera settings table values — are heuristics drawn from drone photogrammetry practice and adjusted for 3DGS’s higher coverage requirements. They are starting points based on field experience, not figures from controlled studies. Calibrate them to your scenes.