OVOW/benchmark/README.md

OVOW-Bench

OVOW-Bench is the evaluation toolkit for instance-level 3D/4D scene reconstruction, and is the benchmark/ component of the OVOW repository. It holds only evaluator code and protocol metadata: it intentionally contains no ground truth, benchmark results, or model predictions. The data is published separately as OVOW_Bench_Data on Hugging Face.

Metrics

The release evaluator reports:

• Scene-IoU AABB and OBB (higher is better)
• Object-IoU AABB and OBB after Hungarian matching and object-count penalty (higher is better)
• Photometric Loss: 100 * mean(normalized RGB MSE) (lower is better)
• N-CLIP: 10 * mean(1 - CLIPSim(pred_render, gt_render)) (lower is better)

Geometry is evaluated after the alignment specified by the protocol. Visual metrics use GT-camera renders of the aligned scenes. The public defaults are in configs/release_protocol.json; the complete definitions are in docs/EVALUATION_PROTOCOL.md.

Repository contents

configs/                 Public evaluator defaults
scripts/                 Evaluation, rendering, validation, and data utilities
tests/                   Unit tests for evaluator helpers and release checks

Benchmark data, generated reports, rendered images, caches, and model predictions are excluded from version control.

Installation

Requirements:

• Python 3.10 or 3.11
• Blender with background Python execution and Eevee or Cycles support
• A local copy or Hugging Face cache entry for openai/clip-vit-base-patch32
• CUDA is optional; CLIP falls back to CPU

The recommended setup uses Conda:

conda env create -f environment.yml
conda activate ovow-bench
python scripts/check_env.py

Conda installs Python and delegates the Python packages to the repository's single source of truth, requirements.txt. Blender remains a separate system dependency.

Alternatively, use Python's built-in virtual environment:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python scripts/check_env.py

The evaluator loads CLIP with local_files_only=True; evaluation does not silently download model weights.

Get the public example data

Download the data with huggingface-cli download (or Git LFS) and pass its local path through --data_root; it can live anywhere writable. The dataset identifier is filled in here once the Hugging Face upload is public:

huggingface-cli download <OVOW_Bench_Data dataset id> --repo-type dataset \
  --local-dir /path/to/OVOW_Bench_Data

The examples below assume the data sits next to the OVOW checkout, so that --data_root ../OVOW_Bench_Data resolves from inside benchmark/. Adjust the path to wherever you unpacked it.

Data and submission layout

Ground truth follows this layout:

OVOW_Bench_Data/gt/<sample_id>/
  scene.blend             # or scene.glb / scene.gltf
  scene.png               # reference image when provided
  metadata.json

Put a method's predictions under a new model key. Do not modify gt/:

OVOW_Bench_Data/<model_name>/<sample_id>/
  scene.blend             # or scene.glb / scene.gltf

For example, predictions for a method named my_method go in OVOW_Bench_Data/my_method/.... User predictions should remain local and must not be pushed to the public data repository.

Validate the included GT subset and checksums:

python scripts/validate_release.py \
  --data_root ../OVOW_Bench_Data \
  --track_manifest ../OVOW_Bench_Data/manifests/example_dynamic.jsonl \
  --check_sha256

Validate a submission before evaluation:

python scripts/validate_release.py \
  --data_root ../OVOW_Bench_Data \
  --track_manifest ../OVOW_Bench_Data/manifests/example_dynamic.jsonl \
  --models my_method

Run the evaluator

python scripts/run_parallel_eval.py \
  --protocol configs/release_protocol.json \
  --data_root ../OVOW_Bench_Data \
  --track_manifest ../OVOW_Bench_Data/manifests/example_dynamic.jsonl \
  --models my_method \
  --gpus 0 \
  --workers_per_gpu 1 \
  --blender_bin blender \
  --out_root /path/to/ovow_bench_outputs

Use a writable output directory outside the repository for large runs. The evaluator is strict by default:

• every requested model must contain every selected sample;
• any failed sample causes a non-zero exit;
• cached reports are reused only when the evaluation configuration and scene file signatures match;
• summaries include sample coverage and refuse incomplete aggregation.

--allow_partial is available for debugging only. Results produced with that flag are marked incomplete and should not be used for method comparison.

Outputs

The output root contains:

results_per_sample.csv
results_summary.json
<sample_id>/run_config.json
<sample_id>/pipeline.log
<sample_id>/reports/final_report.json

results_summary.json records the metric definitions, expected sample count, per-model coverage, and missing rows (if partial mode was explicitly enabled).

Tests

Run the unit tests without writing Python caches:

PYTHONDONTWRITEBYTECODE=1 python -m unittest discover -s tests -v

The unit tests do not require Blender or CLIP weights. A full end-to-end run does.

Scope

This release evaluates already reconstructed scenes. It does not include the OVOW reconstruction pipeline, third-party baseline implementations, historical evaluation outputs, or complete private benchmark data.

Citation

See CITATION.cff at the repository root for the OVOW paper citation.

License

This evaluator is OVOW-authored code and is covered by the repository's MIT license. Note that the assembled OVOW repository as a whole is a mixed-license, non-commercial research release — see NOTICE and THIRD_PARTY.md. The OVOW_Bench_Data dataset is distributed separately and carries its own license.