The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
New ScanNet BAGS outputs
This repository contains the manually refined and metadata-aligned ScanNet BAGS output release.
- 146 scenes
- 3,102 BAGS render frames
- 648 x 484 PNG resolution
- 477 frames removed by the latest manual refinement, including 14 fully removed scenes
Download scannet_bags146_user_refined_aligned_20260809.zip. The archive contains per-scene BAGS PNGs, exact frame ID lists, filtered per-frame metrics, recomputed scene summaries, the global manifest, and QC records.
SHA256: 9e9c9da56237ce697e6e0b12561f6c60f49630034c6cc729dbafb974c1f21a97
Paired-data supplement β 2026-10-10 (partial local recovery)
This supplement does not complete the dataset. The original 146-scene, 3,102-frame BAGS archive above remains unchanged. Surviving local files were matched by exact scene and frame ID against that archive.
| Data recovered | Coverage |
|---|---|
| Original RGB copies (native 1296 Γ 968 PNG) | 5 / 3,102 frames |
| Historical EVSSM intermediate outputs | 8 / 3,102 frames |
| Native ScanNet sensor poses | 247 / 3,102 frames |
| Sensor calibration | 13 / 146 scenes, 52 matrices |
| Sensor depth | 0 / 3,102 frames |
| Complete RGB-D + pose pairs | 0 |
The five RGB frames are scene0005_01/{000006,000008,000010,000024,000030}.
RGB copies preserve native pixels; the BAGS renders are 648 Γ 484. EVSSM outputs
are intermediate predictions, not raw input or independently measured ground truth.
BAGS renders are pseudo-targets. Historical COLMAP variants are included separately
as provenance and have not been verified as the exact final BAGS render cameras.
- Download the partial recovery ZIP
- Supplement README
- All 3,102 frame correspondences
- Missing files by frame
- Coverage and archive validation
- Source paths, roles and checksums
Still missing from this recovery: 3,097 raw RGB images, all 3,102 sensor-depth images, and 2,855 sensor poses. Full source sequences and model checkpoints are not supplied. This is not a complete training-reproduction release. Expected local ScanNet v1/v2 source roots were absent; many processing folders retained metadata but no images. The supplement records the inspected locations and limitations. The original ScanNet terms remain applicable to recovered source data.
Existing code and reconstruction guide β 2026-10-10
Download the 37 MB code bundle | δΈζθ―΄ζ | Detailed reconstruction steps | Actual validation results
The bundle contains 186 existing scripts (22 core pipeline dependencies), local modified BAGS / EVSSM / Mesh2Splat source, the exact final 146-scene / 3,102-frame whitelist, all 146 original export records, 143 surviving model cfg_args files, and available historical queue/QC metadata. It also contains current environment inventories, path configuration, dependency/input preflight and a selected-frame export tool. Unpack the full ZIP; separately displayed helper files require the other bundled source and metadata.
Start with README_ZH.txt / REBUILD_GUIDE.txt. Supply original ScanNet .sens,
scene .txt, cleaned mesh and an explicitly identified EVSSM checkpoint; configure
config.example.json; then use tools/rebuild.py configure, check-env, plan
and finally run. The launcher retains intermediates/checkpoints. Exporting a
missing-frame subset requires explicit --allow-missing; old metrics are never
copied onto regenerated images.
This is an existing-code reconstruction starting point, not an end-to-end validated or pixel-identical reproduction. Current local source snapshots are not proven to be the frozen August revision. The original selected trained BAGS states were not found, and the historical EVSSM checkpoint identity is unverified. Historical runs used 93 scenes at 10k, 38 at 50k, 14 at 100k and one at 89,600 iterations. The 143 cfg_args files are not complete effective training histories. The missing original cfg_args concern scene0024_01, scene0200_00 and scene0452_01. No settings from a different experiment were substituted for those three.
Eight checks passed: shell syntax, fresh path configuration, relocated main help, missing-input planning/refusal, three-environment dependency imports, and strict/ explicitly partial export behavior with actual original ZIP images. All archive entries were checked against their SHA-256 values. No training job or complete GPU reconstruction was run. Consult the guide for data, weights, build, coordinate, selection and metric limitations. In the historical evaluation script, PSNR/SSIM/ LPIPS compare against RAW images, not an independent sharp ground-truth reference.
- Downloads last month
- 32