RuileiG/metric

V1 video evaluation reference package

This dataset contains frozen reference material for the selected SAM3+DINO+Luna v1 evaluation with 13 metrics. It contains 412 source videos / 12,360 task configurations, spanning DyOmni, Move360, ODVista and X360, multiple model input contracts, and 30°, 60°, 90°, 360° yaw schedules.

Move360 0001 and 0344, including all 72 associated configurations, are excluded by the release policy. Original historical artifacts are retained by the authors; they are not part of this release. task_index.json defines the included task IDs and excluded_tasks.json records the exclusions.

Contents

• shards/<video_id>.tar: one independently downloadable shard per retained source video. Tasks share assets within a shard.
• package.json: format and automatic legacy-path mapping. Legacy paths in immutable metadata are provenance, not machine requirements.
• manifests/<video_id>.json: exported file paths, SHA256, sizes and task membership.
• release_policy.json: chosen identity strategy and exclusions.
• release_complete.json: final completion audit. A completed GT calculation is distinct from a finished export; upload the release only after this receipt says COMPLETE.
• validation_report.json: actual portability and consistency validation, with stated limits.

Each extracted task includes actual conditioning inputs; native output geometry and timing; requested camera poses, intrinsic calibration and yaw; input identity anchors and GT eligibility; return/visibility windows; projected GT RGB and masks; frozen CLIP object features; and preparation/integrity records. No generated-video reviews, API credentials or model checkpoint binaries are distributed in the GT shards.

The dataset does not include every original panoramic source frame or the complete raw camera-estimation training/export directories. It includes the prepared target-view reference material needed to evaluate the declared tasks. Rebuilding source annotations is a separate process; scoring requires no access to the authors' raw source directories.

Use

Use the companion evaluation code release metric-v1-portable-20261105. Download the top-level JSON files and the desired video shard, then:

python evaluate.py unpack --dataset /path/to/download --output /path/to/gt \
  --video dyomni__Garden

The unpacker checks member safety and every extracted file hash. It preserves original JSON bytes and resolves paths on load, avoiding machine-local device-number dependencies. Read the companion code's installation instructions for exact SAM3 patches, model revisions and three Python environments.

Metrics

LPIPS; CLIP-Sim_GT; return-window mask IoU; return-window center error; IRR; VRR; content plausibility; yaw MAE; rotation RPE; collapse incidence; unified object-anomaly rate; scene consistency; scene dynamics. Full definitions, Luna rubrics, missingness rules and aggregation are in the code documentation. The original GT delivery files describe eleven GT-related dimensions; the two added scene metrics use generated evidence and do not require new GT assets.

No forward-translation/speed score and no composite score across the 13 metrics are claimed. A missing match penalizes applicable identity/IoU measures according to the frozen protocol; conditional appearance and center metrics require matched masks. NOT_APPLICABLE and technical INCOMPLETE are distinct and must not be silently converted to zero.

Limitations and provenance

GT computation ready means dependencies and integrity checks have passed; it does not certify every automatic object identity as human ground truth. Source RGB and trajectory support can cover only part of a generated timeline; report coverage. The scene-dynamics metric measures intrinsic activity, not general video quality. Luna reviews are model judgments, not human annotations, and fresh requests need not be deterministic.

Original dataset and model licenses remain applicable. This release does not assign a new blanket license to the upstream datasets or model weights. Project authors should provide the final repository citation/license metadata when publishing. The package itself contains no login credential.

Optional offline regression example

examples/x360-hydra90-v1.tar is separate from the benchmark GT shards. It contains one generated Hydra video as native frames, its existing candidate masks and accepted model reviews, a self-contained reference subset, and expected 13-metric results. The generated frames and model reviews are not ground truth labels. It allows the receiving researcher to test the evaluator without a Luna credential or new SAM3/Omega inference. Use the companion tools/score_example.py; all file hashes and numeric/status/coverage comparisons are checked.