Measure the Record. Publish the limits.
A held-out appearance test renders views withheld from training and compares those renders with source images. It measures appearance under the stated camera and scoring conditions, not every property of a Site.
Score views training did not see.
Reserve evaluation views, preserve the split, train on the remaining images, then score renders at the evaluation cameras. TRACE’s Record bundle gate checks distinct views, per-view scores, supporting image comparisons and consistency between documents.
A mean cannot show every local defect. Inspect the renders and record missing coverage, excluded views and any dependence on inherited camera poses.
The warehouse dataset.
The 16 September 2026 warehouse evaluation reports 23.64 dB mean PSNR across 54 panorama views withheld from optimization. Camera poses were inherited and pose-assisted; this is not an independent structure-from-motion test.
The dataset table separates the input inventory, training views and scored views. Three reserved views were excluded because the pinned renderer lacked principal-point support.
The model contains 20 million Gaussians. Visual review did not accept this model as complete; neither the count nor the score establishes complete coverage.
| Measurement | Result | What it establishes |
|---|---|---|
| Input views by source | 2,868 + 1,178 + 149 | Perspective 360 views + drone photos + stills |
| Held-out panorama PSNR, mean | 23.64 dB | Appearance on 54 scored views |
| Registered / training views | 4,195 / 4,138 | Input and optimization inventories |
| Reserved / scored views | 57 / 54 | Three drone views remain unscored |
| Gaussian count | 20,000,000 | Representation size, not accuracy |
Measure compression separately.
The 19 September 2026 delivery contains 20 million Gaussians in a 261,645,206-byte SOG: 13.08 bytes per Gaussian for this export. That is an example, not a fixed format cost.
The separate 31.78–40.64 dB result compares renders of the PLY with renders of its SOG export across 11 frozen views. It measures export fidelity, not reconstruction quality against the real scene.
The public warehouse Record film shows a separate render sequence. These scores are not assigned to that film, and this historical dataset does not certify every path in the current processing service.
Questions, answered.
Why not lead with PSNR?
A single average hides coverage gaps and local defects. We lead with the test method, publish its scope and distinguish image appearance from geometric or metric accuracy.
Is 31.78–40.64 dB the warehouse reconstruction quality?
No. It is PLY-versus-SOG export fidelity. The cited reconstruction appearance result is 23.64 dB over 54 panorama views, with inherited, pose-assisted evaluation cameras.
Is the warehouse evaluation independent of the camera solve?
No. Views were withheld from optimization, but evaluation used inherited camera poses. It does not establish independent structure-from-motion quality, verified scale or survey accuracy.
Where can an agent read the numbers?
The public warehouse receipt summary contains dates, exact values, model identities, source-document hashes and limits. It is a transcription of the original local receipts, not a new benchmark run.
Measurement receiptCapabilities JSONFor agents
Request early access