← sources

DIY iToF + RGB rig: first metric room map and Gaussian splat, still offline

source: reddit (reddit:robotics)published 2026-10-23 03:38ingested 2026-10-24 02:39

retained text

Update on my home-lab RGB-D build: a 320×240 epc660-class iToF engineering sample paired with a Raspberry Pi camera. I am the builder and made the video.



The rig now produces live coloured depth. The mapping pipeline is offline: recover camera poses from recorded RGB frames, use iToF for metric scale, then fuse the depth and train a Gaussian splat in the same coordinates. Multi-view voxel voting removed 27,675 floating points in this capture.



The calibration work mattered more than the rendering. Edge alignment improved from a 2.2 px to a 1.0 px median in our test. Bright-object spill/saturation still needed conditional handling. This is not a general accuracy claim.



Remaining problems: incomplete room coverage, dark walls disappearing, and Pi camera intrinsics that still need a checkerboard check. Real-time Gaussian SLAM is the goal; this result is an offline reconstruction.



EP16 shows the captures, fixes and remaining failures: [https://www.youtube.com/watch?v=pyPj\_dCDZm0](https://www.youtube.com/watch?v=pyPj_dCDZm0)



For a rig like this, would you validate intrinsics and depth scale independently first, or move to joint optimisation once the initial alignment is stable?



Image: AI-assisted episode thumbnail; the episode contains the actual sensor captures.

document fetched 2026-10-24 00:38 · canonical https://reddit.com/r/robotics/comments/1xd3cxx/diy_itof_rgb_rig_first_metric_room_map_and