| # Release scope |
|
|
| This document records what can be run from this snapshot and what requires additional implementation or experiment artifacts. |
|
|
| ## Implemented and directly checkable |
|
|
| - Deterministic synthetic metadata example and generated image/depth files. |
| - Normalized candidate selection and covered-cell metric utilities. |
| - Annotated oracle-gap and basic render-file checks. |
| - Blender/GLB/PLY pipeline sources and native output summarizers. |
| - Author attribution, dataset metadata snapshot and code integrity tools. |
|
|
| ## Important interface differences |
|
|
| `pipelines/run_blend_pipeline.py::select_next_frame` ranks candidates by unoccupied projected pixels minus a fixed overlap penalty. It does not compare candidate probe depths with historical depths in the way required for the paper's depth-disagreement term. Its exported `selected_frames.json` is not the normalized tiny metadata schema. |
|
|
| `scripts/select_views.py` is an independent example based on supplied cell sets and static priors, with an extra 0.15 probe-coverage term. Its sample `oracle_gain` numbers are supplied annotations, not dynamically rendered exact marginal gains. The synthetic image/depth and simplified pose files validate I/O only. |
|
|
| Published Blender pose exports declare camera-to-world rotations and absolute positions; users must reconcile conventions before projection. This release does not relabel or transform already published data. |
|
|
| ## Not included as complete reproductions |
|
|
| - Full matched 900-scene world-voxel baseline sweep and scene-bootstrap confidence intervals. |
| - ERP-LVSM/LoRA training, checkpoints, exact training lists and held-out Replica evaluation. |
| - Full exact-greedy oracle diagnostics and all rebuttal experiment drivers. |
| - Executable TartanGround/OB3D re-encoding workflows and all upstream instance mappings. |
| - Real per-step candidate/selection provenance and full failure-case sequences for the hosted dataset. |
|
|
| Implementing a stub or generating simulated CSVs would not make these experiments reproducible. This repository does not claim that the no-data test verifies paper numbers. Historical H100 inventories record hashes of an earlier local population and are not a public data release manifest or evidence that every artifact is hosted. |
|
|
