Dropbox reconstruction sweep / visual QC
Deep-Reef Reconstruction Atlas
Quality-gated reconstruction candidates from the Dropbox videos, including a navigable full-video atlas and failed global passes retained as diagnostics.
One GPU, full reconstruction stack
The survey atlas and quality-gated photoreal gold scene were trained on one rented RunPod A100 GPU. No multi-GPU cluster was used.
2
global passes
22
trained scenes
18
distinct scenes
4
proxy alternates
4
merge pairs
11
source videos
4656.7 MB
compact splats
Semantics held
Geometry review comes first
Scene index
18 scenes in this view
hero photoreal gold
J2_346.mp4
26:59–28:29 · segment 22
Survey path
1m 30s
180/180
registered
3,708,803
trained Gaussians
4.3
path units
45.7 MB
viewer asset
Claim boundary
Single-GPU arbitrary-scale reconstruction. Promotion depends on the held-out photometric and covariance gates in evidence/quality-gate.json.
Semantic status
Deferred by design. Grounding the promoted scenes later avoids spending the long classification pass on duplicate or rejected geometry.