Splatworks — camera poses (SfM)

ASource

Choose where camera poses come from. Importing is the reliable path; in-browser SfM is experimental and works best with many overlapping, textured photos.

in-browser SfM · experimental

Feature engine

Both built-in engines are pure JS and fully offline. SIFT (gradient-histogram descriptors) matches more reliably on smooth or low-texture scenes than the ORB-style binary descriptors, at a slower detect+match step. OpenCV ORB needs a self-hosted opencv.js next to the app; if it's missing, SfM falls back to the built-in detector. (There is no OpenCV SIFT option — most prebuilt opencv.js binaries ship without SIFT; the built-in DoG+SIFT is the same algorithm and needs nothing.)

Upload photos & run SfM

Take a burst of photos, or record / pick a video and turn it into an image sequence — right in the browser. Needs a camera and a secure (https) page; both feed straight into SfM.

Field of view (horizontal)60°

import poses

COLMAP folder (.txt) Bundle .json

BReconstruction

cameras0

points0

source—

CHand off

Send the poses straight to the trainer — no re-upload — or save a bundle you can load later.

Training resolution

SfM always runs at 512 px; this re-decodes your photos at the chosen size for TRAINING only. The trainer can only be as sharp as its images — 512 px looks soft on a fullscreen viewer. Higher settings cost speed and memory proportionally to the pixels.

Recovered cameras & sparse clouddrag · scroll