Raw DJI .OSV dual-fisheye → Gaussian splat. No stitch.
OSVplat turns a DJI Osmo 360 or Avata 360 .OSV into a flyable 3D Gaussian
splat. It keeps both fisheye lenses as a calibrated rig instead of stitching
first. Copy the clip to your Linux GPU workstation, or hand it to a rented GPU
on Vast.ai or RunPod, pick a preset in a small web UI, and an hour or two later
download .ply, .sog and .spz files.
Live demo: splat.piotrek.uk. Explore an OSVplat splat of Leeds Corn Exchange in your browser, with touch, mouse or keyboard.
A fly-through of a splat trained from a single handheld Osmo 360 walk.
- Reads DJI's raw dual-fisheye
.OSVdirectly. No DJI Studio stitching. It decodes the per-unit lens calibration stored in the file and reconstructs both lenses as a calibrated two-camera rig. That registers about 2× the 3D points of a stitched panorama with lower error (how it works). Stitched 2:1 equirectangular video works too, e.g. a graded D-Log export from DJI Studio or footage from another 360 camera (details). - Complete pipeline, cached per stage. GPU frame decode → sharpest-frame selection → optional person masking → COLMAP SfM on the GPU → LichtFeld Studio training → export. Change a training option and everything before training is reused.
- No GPU of your own needed. The same image runs on a rented Vast.ai or RunPod GPU: the clip goes in and the splats come out over presigned S3 URLs, and you reach the UI over SSH. Stages are sized to the CPUs the container is actually allowed, and GPUs the build cannot run are refused up front (rented GPUs).
- Web queue. Submit jobs, watch progress per stage, see time estimates and PSNR/SSIM, compare runs, run parameter sweeps, and download results. It runs as a systemd service, so jobs survive closing the browser or logging out.
- Removes the operator. Handheld and selfie-stick captures can mask people
(Mask R-CNN, or SAM 3 with text prompts like
"person", "dog"), with an optional review step before the GPU-hour is spent. - Uses the camera's telemetry. Decodes the
.OSVorientation stream (1–4 kHz) and can veto frames whose rotation predicts motion blur. The splat comes out upright, and in metres facing north when the clip has GPS (Avata 360) (how it works).
Low aerial 360 pass along a river. Left: rendered from the splat. Right: the camera frame.
- A Linux workstation with an NVIDIA GPU, RTX 20xx or newer (0.1.4 images run on RTX 20xx and newer; the older 0.1.2/0.1.3 images need RTX 30xx or newer, see troubleshooting #31; developed on RTX 3090s; a Standard run peaked at 10 GB of VRAM), and ~20 GB of free disk per clip. Or a rented one: see docs/cloud.md.
- Clips from a DJI Osmo 360 or Avata 360 (
.OSV), or any stitched equirectangular video (.mp4, 2:1).
With Docker (recommended; needs the NVIDIA Container Toolkit):
git clone https://github.com/pgodlews/OSVplat.git ~/osvplat && cd ~/osvplat
cp .env.example .env && mkdir -p samples data models
sed -i "s/^UID=.*/UID=$(id -u)/; s/^GID=.*/GID=$(id -g)/" .env
docker compose pull && docker compose up -d # or build it here: docker compose up -d --build
docker compose logs queue | grep token # the URL to open
cp /media/$USER/SD/DCIM/DJI_001/CAM_*.OSV samples/docker compose pull fetches the prebuilt image
(ghcr.io/pgodlews/osvplat:latest, currently 0.1.4-rc6); the --build form
builds it here instead, about an hour, once. Details, settings, updating and
verifying the signature: docs/docker.md.
On a rented GPU (Vast.ai, RunPod): start ghcr.io/pgodlews/osvplat:0.1.4-rc6
with your SSH key and two presigned URLs; scripts/presign_s3.py makes them.
Steps, settings and what to watch out for: docs/cloud.md.
Native install (builds the tools on the workstation, about an hour, after the system packages in docs/install.md):
git clone https://github.com/pgodlews/OSVplat.git ~/osvplat && cd ~/osvplat
scripts/setup.sh # pycolmap-cuda, gsplat, LichtFeld Studio
./queue/deploy.sh # installs the queue as a systemd service, prints its URL
cp /media/$USER/SD/DCIM/DJI_001/CAM_*.OSV ~/splat/samples/-
Open the queue's URL (
http://gpu-workstation:8090/?token=…), in a browser on the workstation or any machine on your network. The queue starts paused, so press Resume queue. -
Under New job, pick your clip in Input clip.
.OSVfiles are detected as fisheye rigs automatically. -
Pick a Preset:
- Smoke test: about 10 minutes, on the first 30 s of the clip; checks the whole chain works on your machine. Not meant to be looked at.
- Draft: 15k iterations, 1M splats. A 2.5-minute Osmo clip took 35 minutes on an RTX 3090: a quick look at whether the clip reconstructs.
- Standard: 30k iterations, 3M splats; the settings every measurement in the docs used.
- Max: Standard with full-resolution training and SH degree 3 (more view-dependent colour), for a final result. Slower than Standard.
Trim start/end to skip the take-off, landing or the walk back to the car.
-
Person masking is on by default for Osmo 360 clips (the operator is always in shot) and off for the Avata 360. Change it under Quality flags, where you can also pick the Mask backend: Mask R-CNN (people, no setup) or SAM 3 (anything you name, after a one-time weights download: docs/docker.md).
-
Estimate shows the expected time per stage. Queue submits the job. Click the job name to see per-stage progress, logs and final metrics.
When the job is done, its page shows Download buttons for the .ply,
.sog and .spz files (GET /api/jobs/<id>/files from a script).
Viewing: drag the .sog or .ply into the
SuperSplat editor, or host the .sog with any
static web splat viewer.
Scripting: everything the UI does is a JSON API. See queue/README.md. For example:
TOKEN=$(cat data/.queue_token) # Docker; native: see queue/README.md
curl -H "X-Queue-Token: $TOKEN" -H 'Content-Type: application/json' \
-d '{"config": {"name": "my_clip", "input": {"file": "samples/my_clip.OSV"}}}' \
http://localhost:8090/api/jobsFly or walk low and slow, orbit what you care about, and stay over land. Water, sky, and moving people and cars add nothing and turn the far field into haze. Use a fast shutter, because blur costs more than resolution. For more, see Capturing for a good splat.
| docs/docker.md | Running with Docker Compose, settings, updating, verifying and publishing the image |
| docs/cloud.md | Rented GPUs (Vast.ai, RunPod): SSH, presigned input and output URLs, CPU/GPU discovery |
| docs/install.md | Native install without Docker |
| docs/how-it-works.md | Stages, fisheye rig vs stitching, masking, gyro veto, output formats, with measurements |
| docs/troubleshooting.md | Known traps, numbered (code comments cite them) |
| docs/job-telemetry.md | Per-job timings and machine description: what is recorded, what is not, how to turn it off or upload it |
| queue/README.md | Queue service internals, job options, API, tests |
| docs/osmo360-telemetry.md, docs/avata360-telemetry.md | What is inside a DJI .OSV: lens calibration, IMU, GPS |
- Linux + NVIDIA only (driver 580+). Stages call CUDA builds of ffmpeg, COLMAP and LichtFeld.
- One queue per machine. Multiple GPUs are scheduled, but there is no multi-machine support.
- Insta360
.insvis not supported (its calibration is not decoded). Stitch it to an equirectangular MP4 first. - Long fisheye clips (more than ~500 selected frames) make SfM slow. Trim the clip or raise the sharpness window.
- Tested on two machines so far: a desktop with 2× RTX 3090 and a Ryzen mini PC with an RTX 3090 over OcuLink, both Ubuntu 24.04. Issues and PRs are welcome, especially reports from other GPUs and cameras.
Built on COLMAP / pycolmap (BSD),
LichtFeld Studio (GPL-3.0, built
and run as a separate program), gsplat
(Apache-2.0), FFmpeg, torchvision's Mask R-CNN (BSD),
and optionally Meta's SAM 3 (SAM
licence). Field names for DJI's telemetry come from AdrianEddy's
telemetry-parser.
scripts/colmap_incremental.py is adapted from COLMAP's own example and keeps
its BSD notice.
Not affiliated with or endorsed by DJI.
Copyright © 2026 Piotr Godlewski. Released under the MIT License. Third-party components keep their own licences; see THIRD_PARTY.md.