Ryan-2026/UTok3D-OpenScene-S3DIS-Area5

OpenScene/LSeg on S3DIS Area 5

The reproduction code and reviewer CPU evaluator are available at gfmei/UTok3D-OpenScene. This dataset repository provides the full fused 512-dimensional LSeg features, class scores, predictions, and evaluation records for the OpenScene projection baseline in the revised UTok3D manuscript.

The evaluated split is all 68 S3DIS Area 5 rooms, using a fixed 0.04 m voxel grid and one representative point per occupied voxel. The primary result is 40.2% class-mean mIoU / 49.5% mean class accuracy. Unrounded: 40.16114727433952% / 49.47965993714357%. There are 5,340,122 representatives, of which 4,794,824 have image features (89.788660259% coverage).

Validation is complete for all 68 rooms. The released float32 class scores reproduce every audited prediction and the exact global confusion matrix. Direct CPU matching of the full features against the provided text embeddings differs at one of 5,340,122 representatives because of floating-point operation order: 40.161155177% mIoU / 49.479665235% mAcc, still 40.2% / 49.5%. The complete checks are in reports/score_validation.json and reports/feature_validation.json. Feature release v1.0 pins the generated artifacts. Current reproduction code uses extractor/ on GitHub and on this dataset repository's main branch.

These are a fresh, documented S3DIS adaptation of the official OpenScene LSeg fusion pipeline. They do not recover the unavailable predictions or settings underlying the original manuscript's 40.3% / 37.5% entry.

Files and row alignment

• features/Area_5_*.npy: (N_voxels, 512) full mean fused LSeg embeddings, float32, approximately 11 GB in total. Uncovered rows are zero.
• scores/Area_5_*.npy: (N_voxels, 13) mean class dot products, float32. Per-view score accumulation preserves the original classification arithmetic.
• predictions/Area_5_*.npz: representative_index, prediction, and view_count, all aligned with the feature/score rows. Prediction -1 means no image feature; it is retained as a miss during scoring.
• text_embeddings.npy: (13, 512) normalized text embeddings from the checkpoint's frozen CLIP text encoder, in the class order below.
• room_records/: checksums, source point-file hashes, and comparisons between regenerated features/scores and the audited predictions.
• reports/primary_summary.json and reports/confusion_matrix.csv: primary 0.04 m evaluation; GT rows and prediction columns, with column 13 reserved for no feature.
• feature_generation_config.json and feature_generation_summary.json: complete feature-generation provenance and validation.
• extractor/: standalone evaluation and full inference/fusion scripts.

The first original point in each occupied voxel represents it. Quantization uses floor(float64(original_xyz) * 25), then subtracts the integer room minimum, with the OpenScene FNV64 voxel ordering and augmentation disabled. The prediction and GT label come from the same representative point. There is no semantic majority vote, feature pooling across points, or inverse mapping to all original points. representative_index addresses the exact unaligned Open3DIS preprocess_open3dis/Area_5_*.pth point-file ordering.

The class order is: ceiling, floor, wall, beam, column, window, door, table, chair, sofa, bookcase, board, clutter. mIoU is the mean of class IoUs computed from global confusion counts over all 68 rooms; mAcc is mean class recall from the same counts. It is distinct from ShapeNetPart's instance mIoU_I.

Verify without running LSeg

Install Python 3.10, NumPy, CPU PyTorch, and huggingface_hub. Clone the reproduction code from GitHub:

git clone https://github.com/gfmei/UTok3D-OpenScene.git
cd UTok3D-OpenScene

python extractor/verify_release.py --download --anonymous \
  --release-root /path/to/local/release \
  --gt-root /path/to/s3dis/preprocess_open3dis \
  --mode scores

The script downloads the small score/prediction artifacts and recomputes representatives and all 13-class confusion counts using your local S3DIS GT. It checks point-file hashes, row alignment, file checksums, and agreement with the recorded predictions and global counts. No GPU or LSeg checkpoint is required. --mode predictions directly verifies the saved predictions. --mode counts audits the provided aggregate confusion counts without GT; that mode is an arithmetic check rather than an independent point-label test.

To test the full 512-dimensional features instead of the saved scores:

python extractor/verify_release.py --download --anonymous \
  --release-root /path/to/local/release \
  --gt-root /path/to/s3dis/preprocess_open3dis \
  --mode features

This computes feature/text dot products on CPU. The validation report states any prediction differences caused by floating-point operation order; the original per-view-accumulated scores remain available for exact reproduction.

Regenerate from RGB-D

Use the pinned environment in reports/environment_freeze.txt. A minimal inference setup, with CUDA 12.4-compatible drivers, is:

python -m pip install torch==2.5.1 torchvision==0.20.1 \
  --index-url https://download.pytorch.org/whl/cu124
python -m pip install -r extractor/requirements.txt
python extractor/setup_upstream.py

Obtain the original datasets under their source terms. The exact unaligned point ordering is provided by the S3DIS preprocessing archive linked in Open3DIS DATA.md. Use its preprocess_open3dis directory; do not use its Area 4-only RGB-D example as the Area 5 image source. Obtain both Stanford 2D-3D-S RGB-D archives named area_5a_no_xyz.tar and area_5b_no_xyz.tar. They supply observations for the same 68-room S3DIS Area 5 split. Archive checksums and camera registration records are included in reports/.

Obtain the official LSeg demo_e200.ckpt checkpoint from the upstream LSeg feature-extraction repository. Its SHA256 is f21aeb5532f53a5109d453d7d0d10320701ea11ece9a002291f84f95bda73af4.

python extractor/reproduce.py prepare --gt-root /path/to/preprocess_open3dis \
  --raw-root /path/to/stanford2d3d_no_xyz --manifest /path/to/manifest.json
python extractor/reproduce.py infer --manifest /path/to/manifest.json \
  --checkpoint /path/to/demo_e200.ckpt --output /path/to/point_predictions
python extractor/evaluate_voxels.py --manifest /path/to/manifest.json \
  --predictions /path/to/point_predictions --voxel-size 0.04 \
  --output /path/to/voxel_predictions
python extractor/export_voxel_features.py --manifest /path/to/manifest.json \
  --voxels /path/to/voxel_predictions --checkpoint /path/to/demo_e200.ckpt \
  --output /path/to/regenerated_features

Inference uses 3,547 perspective views (every fifth sorted view per room), image long side 320, LSeg base/crop 640, single scale with horizontal flip, relative depth tolerance 0.25, a 10-pixel border, and text prompt a [class] in a scene. Area 5b camera poses are registered into the S3DIS world frame. No semantic GT labels enter feature generation or inference. Network parameters are loaded strictly from the official checkpoint. CUDA float16 image inference and a validated SDPA attention implementation are recorded in the provenance; feature and score sums are float32.

Sources

Original RGB-D scans, point coordinates, point-level ground-truth labels, and model checkpoints are not redistributed here. Acquire those from their upstream sources. Dataset and model use remains subject to the original source terms. Cite OpenScene (Peng et al., CVPR 2023), LSeg (Li et al., ICLR 2022), S3DIS (Armeni et al., CVPR 2016), and Stanford 2D-3D-S (Armeni et al., 2017) when using this reproduction. Source commits are recorded in reports/provenance.json.