Back to project page

Underwater 3DGS Demo Track

This page is a public-facing research track, not a claim that aqua_localization already ships a production 3DGS mapper. It frames the next experiment: convert synchronized underwater camera frames, estimated poses, depth, and MBES structure into a reproducible splat export.

GitHub Pages ready static HTML ROS bag inputs v0.3 sample pack

Viewer Controls

Splat density Water attenuation

Current Artifact

Generate a JSON manifest for camera frames, trajectory, depth priors, and optional MBES/submap inputs.
Create a scaffolded dataset pack with images/, depth/, mbes/, and transforms_stub.json.
Extract real ROS bag image frames into images/ and index them in frames.json.
Match extracted frames to nearest odometry and write transforms.json with CameraInfo or manual intrinsics in aqua or nerfstudio format.
Keep localization metrics separate from reconstruction quality claims.

Play Sample Viewer

The v0.3 Release includes a 20-frame Tank Dataset nerfstudio input pack: images, matched /apriltag_slam/GT poses, manual stereo intrinsics, and zero skipped transforms. Open the viewer to play the frames with the matched 3D camera path before downloading the zip.

Play sample viewer Open training smoke guide Download sample pack

Manifest To Transforms

The reproducible track now reaches `images/`, `frames.json`, nearest-neighbour odometry poses, CameraInfo intrinsics, and base-to-camera extrinsics in `transforms.json`. Use the nerfstudio format when handing the pack to external reconstruction tools.

ros2 run aqua_localization export_3dgs_pack_pipeline.py \
  --bag datasets/public/tank_dataset/short_test_ros2 \
  --dataset "Tank Dataset" \
  --sequence short_test \
  --out /tmp/tank_short_test_3dgs_pack \
  --max-frames 120 \
  --stride 2 \
  --max-time-diff 0.05 \
  --base-from-camera -0.25 -0.45 0.0 0.0 0.0 0.0 1.0 \
  --camera-intrinsics 612 512 655.0 655.0 306.0 256.0 \
  --trajectory-topic /apriltag_slam/GT \
  --format nerfstudio

Open sample pack workflow

Why It Helps

Stars come faster when visitors can see the ocean reconstruction idea immediately. The engineering bar stays honest by labeling this as a visualization and dataset-packaging track until a real 3DGS training/export pipeline lands.

1400synthetic preview splats

3candidate input streams

0benchmark claims on this page