RB-1903 / J2-1128 / Dropbox source

Continuous Deep-Reef Mission

A 90-second ROV traverse recovered as a connected camera path, a trainable 3D scene, and a time-addressable evidence record.

180/180
poses registered
165,885
trained Gaussians
1 GPU
local processing
Motion becomes spatial evidence

Play the video. Watch the recovered camera move.

Sequential feature matching establishes local motion across 90 seconds sampled at 2 fps. The solve retains long power-of-two baselines while excluding three ambiguous retrieval edges from the repetitive coral scene. Every highlighted position below is a registered video frame, not an illustrative tour path.

Pose-synchronized mission
Video time drives the recovered camera position
0:00.0 / 0:00.0
J2-1128 / RB-1903 2.0 pose samples/s
Loading recovered trajectory…
Registered poses
— / —
Sparse points
Verified graph edges
Current pose
Claim boundary

This is a real monocular camera solution and reconstruction. Its shape and continuity come from the video; its absolute scale and global navigation remain unresolved until USBL, DVL, altimeter, or another measured reference is fused.

Recovered and trained entirely on one consumer workstation using CPU feature matching and an Apple Silicon GPU for Gaussian splat optimization.

Trained scene

The same pose chain becomes a Gaussian splat.

Drag to orbit, scroll to zoom. A static source frame remains available when WebGL is disabled.

Semantic motion proof

Every frame in the opening 45 seconds is classified.

The existing dense Falcon pass uses one stable class palette across 180 frames, so coral, sediment, rubble, fish, and water-column evidence stay comparable while the vehicle moves.

Loading semantic motion record…