3DGS workflow turns a 10-minute OSMO360 walk into a dense reconstruction on RTX 4090
janusch_patas · x · 2026-07-29
A 3DGS workflow demo shows how to turn a 10-minute DJI OSMO360 walk into a dense reconstruction using seam masks and a 15M-point training setup.
- Capture: a 10-minute OSMO360 walk
- Preprocess: frame extraction plus SAM3 and seam masking, producing 774 images
- Reconstruction: spherical SfM in Metashape using the mask
- Training: cubemap conversion, then 3DGS training in LichtFeld Studio with 15M, SH2, MRNF, PPISP on an RTX 4090 for about 10 hours
- The author says the seam mask reduces errors around the stitch line of dual-fisheye footage
More from Multimodal
- SenseTime open-sources SenseNova-Vision, a unified model for detection to 3D reconstruction — socialwithaayan · 2026-07-29
- ChatGPT demo shows MCP driving a 3D avatar with custom animations — BLUECOW009 · 2026-07-29
- Reddit user shares a Seedance 2 western-town video demo — Far-Teaching3402 · 2026-07-29
- Applied AI Cohort 5 opens with 24 prototypes and 7 production system breakdowns — arpit_bhayani · 2026-07-29
- Seedance 2 turns a moon-knight prompt into a cinematic magic sequence — LudovicCreator · 2026-07-29
- Kimi K3 faces Claude Opus 5 in a one-image 3D world-building test — Prompt Engineering · 2026-07-29