Concept3D4D — compact 2k

Private internal evaluation. 639 static images + 1,361 video frames; 828 queries and 5,106 query/frame pairs.

The package contains original-resolution RGB, fixed query manifests, SAM3 proposals and masks, supplied instance masks, prepared 3D annotations, and the frozen annotation-supported scorer. The original query IDs, image bytes, proposal keys and GT are unchanged. Historical absolute paths are retained as identity strings and resolved through asset_registry.json; those directories are not required.

Download

Access to this private repository is required. Log in with an authorized Hugging Face account and pin the dataset commit shown under Files and versions:

hf auth login
hf download qicq1c/Concept3D4D --repo-type dataset \
  --revision <dataset-commit> --local-dir Concept3D4D
python Concept3D4D/prepare.py \
  --archive Concept3D4D/compact2k-v1.tar.gz --output ./concept3d4d-data

prepare.py checks the archive SHA-256, rejects links and unsafe paths, and verifies every extracted file. It will not overwrite an existing output directory. Allow space for both the archive and extracted data.

Verify and replay the reference score

Use Python 3.12. Install CPU scoring dependencies in a separate environment if the model environment has incompatible pins:

python -m pip install -r concept3d4d-data/requirements-score.txt
export PYTHONPATH="$PWD/concept3d4d-data/code${PYTHONPATH:+:$PYTHONPATH}"
python -m scripts.concept3d4d.portable_runtime verify \
  --dataset ./concept3d4d-data
python -m scripts.concept3d4d.portable_runtime score \
  --dataset ./concept3d4d-data --output ./concept3d4d-reference --reference

The reference replay uses the saved Small E280 predictions. It compares all overall, category and source metrics, plus the frame-event hash, against the original result. It does not run model inference.

Release checks: file hashes and frozen populations verified; 27 regression tests passed; reference metrics and frame-event hash reproduced exactly; a separate Small E280 single-query GPU inference smoke test passed. This smoke test is not a new full benchmark result.

Evaluate a Full or Small checkpoint

Use the corresponding SpatialEncoder repository and its installed inference environment. Keep the model's configured input resolutions; 518 below is the Small 518 example.

export PYTHONPATH="$PWD/concept3d4d-data/code${PYTHONPATH:+:$PYTHONPATH}"
CUDA_VISIBLE_DEVICES=0 python -m scripts.concept3d4d.portable_runtime infer \
  --dataset ./concept3d4d-data --output ./concept3d4d-my-model \
  --model-repo /path/to/spatial_encoder_v2_small \
  --checkpoint /path/to/checkpoint.pt --profile lite \
  --resolution 518 --spatial-resolution 518
python -m scripts.concept3d4d.portable_runtime score \
  --dataset ./concept3d4d-data --output ./concept3d4d-my-model

For Full, use its repository, --profile full, and the checkpoint's resolutions. For multiple GPUs, run one process per GPU with distinct --shard 0 ... N-1 and common --shards N, using the same output directory. Score only after every process succeeds. The scorer rejects missing shards, stale fingerprints and incomplete query/frame outputs.

Results: concept3d4d-my-model/annotation_supported_v1/scores/ours/summary.json. ours is the adapter's output key, not a claim that the checkpoint is the reference model.

Protocol

The default protocol matches the existing concept3d4d_compact2k_20261010/image_by_category.md and video_by_category.md tables. Those tables already use concept3d4d_annotation_supported_v1; this package does not substitute the native-source protocol or introduce a new scoring rule. Their reference copies are in reference/annotation_supported_v1/.

The frozen annotation-supported population contains 2,271 queried frame/category pairs, not 2,271 scored positive pairs:

Table	Queried pairs	Evaluated pairs with eligible GT	GT boxes
Image by category	636	636	929
Video by category	1,635	453	1,292

The remaining zero-GT pairs are skipped, not counted as negatives. The other 2,835 inference pairs belong to the existing separate protocols: nuScenes tracking, KITTI-360 window AP, ROPE instance regression, and the AstroVision instance pilot. All 5,106 inference inputs are preserved; native-only results are not pooled into these category AP tables. For example, the image table's celestial category remains N/A, rather than importing an AstroVision instance-regression score.

• concept3d4d_annotation_supported_v1: confidence-ranked, one-to-one matching within each frame/category query.
• Candidate matching requires 2D IoU > 0.3 with an eligible same-query GT box. Unsupported predictions are ignored; queries without eligible GT are not treated as negatives.
• AP uses a 101-point precision envelope and 3D IoU thresholds 0.05–0.50. JSON AP values are fractions, not percentages.
• Metric-scale, WildBox relative-scale, and annotation-pilot results remain separate. WildBox normalization uses predictions only; no GT scale fitting.
• ROPE uses supplied instance masks/boxes; AstroVision uses supplied target boxes. These are not unrestricted text-detection settings.
• Video geometry is evaluated per frame with online 2D association; this does not establish temporally accurate 4D tracking performance.

This portable runner reproduces the annotation-supported protocol, not each source's native leaderboard scorer. Prepared native-only GT/calibration and historical native-metric tables are included. External devkits and their separate runtime environments are not bundled; native launchers still require path/environment setup. The existing native evaluation commands should not be pointed directly at this portable package.

Selection SHA-256: 7ece9a118dd2cea8b6e45cf33267b013acdef9f703f51ddb85b33bc73d0f923a
GT frames SHA-256: 64c8308a85d96f3311e5af868fde31592938db362c00f54cd680b7429943aa0a

Contents

Directory	Contents
assets/	Content-addressed original RGB and supplied masks
selection/	Frozen queries and selection record
proposals/	Fixed detections, masks and provenance
annotation_supported_v1/	Prepared GT and scoring protocol
wildbox/	Query mapping for prediction-only scale normalization
native/	Prepared native GT/calibration; original nuScenes metadata tables and maps (not full-source RGB)
code/	Frozen geometry adapter/scorer and portable path wrapper
reference/	Small E280 predictions; existing comparison tables
package.json	Population, file sizes, hashes and provenance

Sources include the selected static WildDet3D/Omni3D-derived subset, animal extension, Objectron, ROPE, WildBox, nuScenes, KITTI-360, AV2, and two explicitly labeled annotation pilots. Exact media and source membership are in reference/tables/media_index.csv and source_counts.csv.

Access and source terms

This is an access-controlled internal evaluation copy, not a public dataset release or a grant to redistribute upstream content. Original source-specific licenses, access agreements and attribution requirements remain applicable. No blanket Apache-2.0 license is assigned to the mixed-source images and annotations. The repository owner has stated that redistribution authorization exists; detailed source-by-source scope must be documented before any public release. Keep this repository private and grant access only to authorized collaborators.