Capture Tutorial For 3D Gaussian Splatting

Object, room, street, landscape, and drone captures need different movement patterns. A small object benefits from multiple orbits at different heights. A room needs wall passes, corner passes, and coverage of floors, ceilings, and occluded furniture sides. A large outdoor scene needs broad baseline variation and often oblique views rather than only front-facing photos.

Before shooting, mark what must survive in the final splat. If the final viewer needs a clean front facade, put more images around edges, windows, vegetation, and reflective areas. If the final output is a walkable room, spend extra time on doorways, corners, glossy floors, under tables, and transition spaces between rooms.

Still photos are usually the safest choice because they are sharper, higher resolution, and easier to inspect. Video can be faster, but it often introduces rolling shutter, compression, motion blur, and too many near-duplicate frames. If you use video, record at high frame rate, move slowly, and extract frames at a rate that keeps overlap without flooding COLMAP with redundant images.

For video extraction, do not simply take every frame. You want frames that differ enough to add parallax but still overlap heavily with neighbors. A practical starting point is one to three frames per second for a slow walk, then reduce or increase based on scene size and motion. When the camera turns quickly, discard frames from the turn because they often contain blur.

The alignment stage wants stable visual features; the training stage wants consistent colors. Lock exposure and white balance when you can. If the camera changes exposure every few frames, the trainer may bake the lighting change into Gaussians and produce flicker, color bands, or floaters.

Soft, even lighting is usually better than dramatic lighting. Harsh shadows move as you walk and can become false texture. Reflections are not always fatal for 3DGS, but mirror-like surfaces still need many angles and stable surroundings. For small objects, remove cluttered backgrounds or use masks later if the background dominates feature matching.

The golden rule is overlap plus parallax. Overlap means two neighboring images see the same features. Parallax means the camera moved enough that the alignment system can recover depth. Rotating in place gives overlap but weak parallax; walking too far between shots gives parallax but poor overlap. Good capture lives between those extremes.

For objects, start with a base orbit where the full subject stays in frame. Then add a higher orbit, a lower orbit, and close detail passes. For rooms, walk the perimeter first, then add a center pass and point the camera into corners and behind furniture. For buildings, combine near detail passes with wider context views so COLMAP can connect local details into one global model.

Apps such as Scaniverse, Polycam, KIRI Engine, Luma AI, and RealityScan can hide parts of the pipeline, but the same capture physics still apply. Slow motion, complete coverage, and stable light matter more than the brand of app. If the app can export PLY, SPLAT, SPZ, or raw posed images, keep the highest-quality source as your archive.

360 cameras are useful for rooms and fast coverage, but equirectangular images often need special handling. Some pipelines split panoramas into rectilinear views; others can process equirectangular images directly. Avoid putting the tripod, operator, or black camera nadir region in the final field of view, or be ready to mask/crop those areas before alignment.

Always do a field review before packing up. Look for motion blur, missing backsides, unreadable dark areas, and areas only visible from one angle. If you see a hole in the capture path, shoot a small bridge sequence from a known area into the missing region.

The most expensive capture mistake is discovering missing coverage after the location is unavailable. A five-minute validation pass can save hours of alignment and training. When in doubt, shoot extra views with stable light instead of hoping the trainer will invent the missing detail.