Camera Settings & On-Site Operation

The 5 Iron Rules (Universal)

Regardless of device, these 5 rules are non-negotiable. Burn them into muscle memory before worrying about device-specific differences.

Lighting Condition Priority

3DGS loves uniform diffuse light — its worst enemy is "the specular highlight moved between angles." Inconsistent lighting directly causes color discontinuities across viewing angles in the trained model.

Rule of thumb: Overcast skies are God's gift to 3DGS shooters. If you can choose the date, always choose overcast.

Video Frame Extraction vs Photo Burst

The industry is split on this. Our judgment: beginners use video, professionals use photos.

Device-Specific Parameters

Below is an excerpt from our studio's capture manual, providing "power-on-and-go" parameter presets by device.

Phone (iPhone / Android Flagship)

iPhone note: The system camera's "Auto HDR" actually varies exposure strategy per frame, causing brightness inconsistency between adjacent frames. Use Halide's Manual HDR mode, or disable system HDR entirely and use fixed exposure.

Mirrorless / DSLR

Why turn off stabilization? In-body stabilization (IBIS) compensates shake by physically moving the sensor — meaning each frame's sensor position differs slightly. SfM assumes the sensor is fixed on the lens optical axis. IS breaks this assumption. On a tripod, it's even worse: the IS system "can't find shake" and creates micro-vibrations of its own.

360° Camera (Insta360 X5 / Theta Z1)

360° cameras are efficiency powerhouses for large spaces. In April 2026, Insta360 partnered with Splatica to offer automated 360° video → 3DGS pipelines: 5-minute capture → model ready in hours.

Critical workflow: .insv → JPEG export

Never feed raw .insv files directly to training software. Export through Insta360 Studio first:

1. Open Insta360 Studio → select .insv files

2. Export settings: maximum resolution (X5 reaches 8192×4096)

3. Disable FlowState stabilization (stabilization alters inter-frame geometry)

4. Enable Direction Lock

5. Enable PureShot / Clarity Enhancement

6. Export as JPEG (or DNG for color grading)

Training failure rate with raw .insv exceeds 80%. After export, use a CubeMap split script to cut into 6 faces, then feed to COLMAP.

Action Camera (GoPro Hero 13 / Insta360 Ace Pro 2 / DJI Action 5)

Action cameras are the best value for outdoor work. The triple-mount rig — three cameras on one pole, voice-synced trigger — captures museums and galleries an order of magnitude faster.

Triple-mount details: A 2m carbon fiber pole with GoPros at top, middle, and bottom. Use GoPro Labs firmware "voice trigger" for synchronized start. Three cameras recording simultaneously = continuous three-height coverage in a single pass.

Drone (DJI Mini 4 Pro / Mavic 3 Pro / Matrice 4E)

DJI Terra V5.2 workflow: Since July 2025, DJI Terra natively supports 3DGS output. Flight planning → autonomous flight → photo transfer → one-click .ply generation. For building exterior projects, this is currently the most hands-off end-to-end solution.

Small Spaces: The "Hard Nut" — No Room to Back Up

Small spaces (10–30㎡, like a single room) are 3DGS's hard nut. The reasons: you're only 1m from the wall, the frame is filled with featureless white wall; no continuous objects provide feature points; lens field of view isn't wide enough, each frame covers too little.

The solution: three-layer shooting + loop path.

Three-Layer Shooting

Switch between three heights to ensure no vertical information gaps:

Transitions between layers still need 5–6 gradual frames — never jump heights abruptly.

Loop Path Pattern

┌─────────────────────┐ │ ←──── Start ──────
│ │ ↑
↓
│
Outer loop: walk along walls │ │
┌───────┐
│
│ │ │
│
←─→
│
│
│
Inner loop: orbit central object │ │
│
↑
│
│
│ │ │
└───────┘
│
│
Key objects: extra 3 orbits │ ↓
↑
│ │
───── End ─────→
│ └─────────────────────┘

Outer loop follows walls (facing opposite wall), inner loop orbits central object (facing inward), key objects get 3 extra orbits (high/mid/low).

Three Traps in Small Spaces

Trap 1: Macro trigger. Phone cameras switch to macro mode when <30cm from an object — only a small area is sharp, background completely blurred. Algorithm fails immediately. Maintain 30–50cm safe distance; step back and use 0.5× ultrawide rather than getting close.

Trap 2: Rotating in place. Standing against a wall and spinning 360° ≠ full coverage. The algorithm sees only "a panorama from one point" with zero depth. You must have a translational trajectory.

Trap 3: Featureless white walls. 3DGS's worst enemy. Three rescue techniques:

• Step back: Include ceiling, floor, and wall in one frame — give the algorithm 3 reference planes

• Shoot at an angle: Photograph walls obliquely so edge lines enter the frame

• Add texture: In your own space, hang a painting or poster as a feature anchor point

Large Spaces: The Opposite Problem — Insufficient Coverage

Large spaces (buildings, campuses, city blocks) have the opposite problem — you can never cover everything in one pass. 500 shots sounds like a lot, but spread across all facades of a three-story building, each angle gets only 2–3 shots — far from enough.

The solution: grid segmentation + multiple flights.

Grid Segmentation Method

1. Divide the large space into 30×30m grid cells

2. Each cell gets a complete three-layer pass (ground 3 layers + drone 3 layers)

3. Leave 30% overlap between cells for SfM stitching

4. Each cell: estimated 80–150 shots (ground) + 50–100 shots (drone)

A 100×100m building complex requires approximately 9 cells × 150 = 1350 ground shots + 9 × 80 = 720 drone shots, totaling ~2000 images. Complete over 2–3 days (same time window each day to ensure lighting consistency).

Handling Moving Objects

3DGS assumes the scene is static. Anything that moves between two frames becomes a chaotic mess of floater artifacts.

Next Steps

• Finished shooting, how to organize files → Enter 04-Dataset Organization & Archiving

• Shot video, how to extract frames → Enter 05-Video Frame Extraction & Quality Filtering

• Photo colors inconsistent → Enter 06-Color Grading & White Balance Unification

• ← Previous chapter: 02-Site Recon & Shot Planning