Documentation
Capture guide for COLMAP
How to film and prepare indoor showflat walkthroughs so COLMAP registers most frames and builds a dense sparse model.
VirtualBhoomi scores your burst against an indoor benchmark. Registration and 3D point density drive most of the score; resolution and exposure matter but rarely fix a failed walk on their own.
1. Motion and overlap (registration)
Goal: Most frames register in COLMAP (registration rate ≥ ~0.8; aim for ~95%+).
| Do this | Why |
|---|---|
| Walk slowly | Fast motion causes blur and large baselines between frames; feature matching breaks down. |
| Keep 70–85% overlap between consecutive views | Too little overlap means views do not share enough features; standing still adds redundancy without new viewpoints. |
| Add doorway bridging shots | Doorways and narrow openings are weak links; pause for 2–4 extra frames looking through each door before and after crossing. |
| Smooth path, no sudden pans | Prefer a steady arc along walls; avoid whip turns and vertical-only spins without lateral motion. |
| One consistent height per room, slight variation across the burst | Stable height helps overlap; small up/down shifts add parallax (see texture section). |
| Drop blurred and duplicate frames before COLMAP | Near-identical or soft frames inflate image count without helping reconstruction and can confuse matching. |
Quick check: If only a handful of images register out of a large burst, fix capture and curation first—do not rely on higher resolution alone.
2. Sparse reconstruction and texture
Goal: Enough 3D points per registered image (≥ ~300; benchmark reference ~390).
| Do this | Why |
|---|---|
| Add furniture, art, rugs, plants | Bare rooms give few stable features; textured surfaces improve matching and triangulation. |
| Avoid long stretches of bare white walls | Low texture means few tracks; angle slightly toward corners and details when possible. |
| Vary height slightly | Small shifts (e.g. chest vs eye level) add viewpoint diversity without breaking overlap on the same path. |
| Keep sharp focus across the burst | Soft focus removes fine detail needed for feature matchers. |
| Include floor–wall boundaries and fixtures | Edges and patterns (tiles, cabinetry) help COLMAP tie views together. |
| Minimize rolling shutter and heavy stabilization crop | Use a stable hold or gimbal; avoid changing digital zoom mid-walk. |
Quick check: Very few 3D points or ~0 points per registered image means the scene lacked matchable structure—re-shoot with more texture and overlap before tuning COLMAP settings.
3. Resolution
Goal: Enough pixel detail for reliable features (prefer ≥12 MP for showflat stills when possible).
| Do this | Why |
|---|---|
| Use full-resolution stills or video at native sensor resolution | Heavy in-camera downscaling (e.g. ~2 MP) limits detail; registration can still fail at any resolution. |
| Avoid unnecessary 4K→720p pipelines | Export frames for COLMAP at the highest practical resolution your device and pipeline support. |
| Consistent aspect ratio and focal length | Use one camera model per burst; do not mix wide and telephoto clips. |
4. Exposure and lighting
Goal: Mean luminance in a usable band (~0.45–0.55); preserve wall and floor texture for matching.
| Do this | Why |
|---|---|
| Reduce exposure on bright facades | Blown highlights remove texture; quality scoring may flag over-exposed bursts. |
| Close blinds or diffuse direct sun | Windows and glossy surfaces cause clipping and unstable features. |
| Avoid an extreme HDR “flat” look | Walls need visible grain or pattern for matchers, not uniform gray or white. |
| Keep lighting stable during the walk | Flickering LEDs or auto-exposure pumping change appearance between adjacent frames. |
5. Pre-COLMAP checklist
- Slow walk with 70–85% overlap between neighbors
- Bridging shots at every doorway and tight corridor
- Blurry and near-duplicate frames removed
- Room has visible texture (not empty white boxes)
- Focus sharp; exposure controlled (no blown windows)
- Frames exported at full practical resolution
- Single camera / consistent intrinsics for the whole burst
6. Priority when fixing a poor score
| # | Symptom | Action |
|---|---|---|
| 1 | Low registration rate | Slow walk, overlap, doorway bridges, cull blur/duplicates |
| 2 | Low points per registered image | Add texture, vary height slightly, sharp focus |
| 3 | Over-exposure advice in quality report | Reduce exposure, close blinds, preserve wall/floor detail |
| 4 | Resolution advice in quality report | Full-res stills (≥12 MP preferred for showflats) |
7. Reference metrics
Typical decent indoor burst targets (for comparison when you open a COLMAP quality report in Studio):
| Metric | Target band (approx.) |
|---|---|
| registration_rate | ≥ 0.88 (excellent ≥ 0.95) |
| points_per_registered_image | ≥ 300 (excellent ~390+) |
| megapixels_mean_sample | Higher is better; low values trigger resolution advice |
| mean_luminance_01_sample | ~0.45–0.55; too high → over-exposure |
Re-run COLMAP quality after changing capture or frame curation. Large score gains usually require better registration and point density—not resolution alone.