HaofanCao/deja-cue-data

Déjà Cue Fixed-Feature Evaluation Assets

Frozen features, fixed evaluation settings, exact reference windows, and statistical results for Déjà Cue.

Haofan Cao · Zhichao You · Yunkai Yang · Liang Guo · Jie Wang · Chongshou Li

📄 Abstract

Tracking links observations of the same object through visual change, yet cannot by itself determine when the object is empty or filled, intact or cut. Déjà Cue formulates identity-conditioned state-moment retrieval: given a tracked-object history and alternative state descriptions, it localizes an interval in which each described state holds. The method turns the alternatives into a vocabulary-relative coordinate system, subtracts their state-balanced centroid from each description, calibrates frame scores, and scans multiple durations within contiguous visible runs using a frozen encoder. On 78 VOST histories, changing only the query reference nearly doubles R@1 at tIoU 0.5 from 10.3% to 20.5% and raises Top-1 tIoU from 16.0% to 21.5%.

🔥 News

• 2026.08.05 · The official code and fixed-feature evaluation assets are now open source.

📦 Dataset Summary

This repository contains the same data/ directory as the companion code release. It reproduces retrieval, interval selection, metrics, and statistical calculations without redistributing native video or rerunning the frozen encoder.

Dataset page: HaofanCao/deja-cue-data

Group	Contents
VOST evaluation	78 histories, 156 states, 312 descriptions, reference intervals, and frozen features
Prompt variants	Raw, photo, definite, and normalized three-form text features
Development set	5 histories, 13 states, 26 descriptions, 46 positive episodes, and 86 training records
Seven-history evaluation	7 histories, 16 states, 32 descriptions, 59 reference episodes, and 8 auxiliary tracks
Reference results	Exact selected windows and the rows used for the reported statistical calculations
Evaluation settings	Fixed VOST split lists, final cohort information, method settings, and checksums

The dataset contains 288 files under data/. DATA_MANIFEST.json records the byte size and SHA-256 digest of every one.

The Hub's n<1K label describes the number of dataset examples, not storage size. The complete repository occupies about 51.5 MB.

⬇️ Download and Verify

The complete dataset is published at HaofanCao/deja-cue-data. With the current Hugging Face CLI, download it with:

export DEJA_CUE_DATASET_ID=HaofanCao/deja-cue-data
hf download "$DEJA_CUE_DATASET_ID" --type dataset --local-dir deja-cue-data
cd deja-cue-data
python verify_data.py

The script checks every file name, byte size, and SHA-256 value. A Python download is equivalent:

from pathlib import Path

from huggingface_hub import snapshot_download

dataset_root = Path(
    snapshot_download(
        repo_id="HaofanCao/deja-cue-data",
        repo_type="dataset",
        local_dir="deja-cue-data",
    )
)

🔎 Load One History

import numpy as np

visual_path = dataset_root / "data/features/siglip2/H001/visual_features.npz"
with np.load(visual_path, allow_pickle=False) as arrays:
    frame_indices = arrays["frame_indices"]
    visual_features = arrays["visual_features"]
    visibility_count = arrays["visibility_count"]

print(frame_indices.shape, visual_features.shape, visibility_count.shape)

Primary visual archives contain:

Field	Meaning
frame_indices	Original temporal indices for observed target frames
visual_features	Unit-normalized 768-dimensional SigLIP 2 features
visibility_count	Number of target-lineage masks contributing at each frame

Primary text archives contain state_ids, state_texts, and unit-normalized text_features. Prompt-variant archives additionally contain variant_names. JSON manifests define sibling-state grouping, inclusive reference intervals, source-component aggregation, and relative feature paths.

🔗 Use with the Code Release

The companion GitHub repository already includes the same data/ files, so a complete clone can reproduce the CPU results immediately:

git clone https://github.com/HaofanCao/DejaCue.git
cd DejaCue
python -m pip install -e ".[test]"
python scripts/reproduce_main.py --device cpu

When using a separate Hub download, place or link its data/ directory at the code repository root without changing the relative paths. The code repository supplies the data loaders, retrieval implementation, reproduction commands, annotation and feature tools, and tests.

🗂️ Repository Layout

data/
  benchmark.json              Primary benchmark manifest
  features/                   VOST visual, text, and prompt features
  learned/development/        Frozen learned-decoder development inputs
  seven_history/              Seven-history features and auxiliary tracks
  reference/                  Exact windows and statistical results
  protocol/vost/              Fixed VOST split lists and protocol notes
DATA_MANIFEST.json            File sizes and SHA-256 checksums
DATA_LICENSE.md               Data and derived-feature terms
verify_data.py                File verification script

Evaluation identifiers such as H001 are stable IDs used to align features, states, descriptions, and references. All manifest paths are relative to the repository root.

🌐 Data Sources

Asset family	Upstream research source	What this repository contains
Primary histories	VOST	Derived object-local features, evaluation IDs, fixed settings, and reference intervals
Six seven-history sequences	HyperNeRF	Derived target and auxiliary-track features plus evaluation metadata
Coffee-martini sequence	Neural 3D Video	Derived target features and evaluation metadata
Recurrence diagnostic	D-NeRF	Predictions and reference intervals, not rendered RGB
Feature encoder	SigLIP 2	Unit-normalized derived embeddings; no model weights

Feature extraction uses google/siglip2-base-patch16-224 at revision 75de2d55ec2d0b4efc50b3e9ad70dba96a7b2fa2. Upstream media and model terms continue to apply to derived assets; see DATA_LICENSE.md.

📝 Annotation and Included Files

For VOST, two state descriptions per state are fixed before labeling. Two annotators independently label target-lineage frames as pre-state, transition, post-state, or unobserved while blind to method scores, and every disagreement is resolved to consensus. The 100-history selection retains 78 qualifying histories and excludes 22 with no qualifying event.

The dataset includes the split lists, final cohort information, annotation validation scripts, and agreement and annotation-sensitivity summaries. It does not redistribute native RGB, target-lineage masks, pretrained encoder weights, generated decoder checkpoints, the two complete annotation files, or the resolved framewise consensus.

✅ Intended Uses

• Reproduction of the primary and extended VOST evaluations.
• Reproduction of the seven-history coordinate study.
• Training the documented eight learned decoders on the five-history development set and evaluating generated checkpoints.
• Recalculation of the reported robustness, diagnostic, and statistical summaries.
• Comparison of alternative retrieval methods under the same frozen representation and evaluation protocol.

⚠️ Limitations

The task assumes target-lineage masks, a closed sibling vocabulary, and one selected interval per description. It has no calibrated null output for an absent state. VOST uses binary event-adjacent states and scores one designated occurrence; the dataset also contains 13 additional pre-state and 20 additional post-state occurrences. The seven-history set adds multi-state, interior, recurrent, and progressive cases, but remains small and its three independently held-out histories show mixed coordinate effects.

Fixed features isolate retrieval logic from encoder execution and do not establish behavior for other encoders. Tracking stresses use controlled feature mixing and identity switches rather than pixel-level mask drift. The release does not establish performance for automatically discovered identities, open vocabularies, broad multi-state collections, every natural tracking failure, or hardware-specific latency.

🛡️ Privacy and Responsible Use

No native RGB, video, masks, direct identifiers, or model weights are included. The embeddings remain derived from external research media and must be handled under the corresponding source terms. These assets are intended for scientific retrieval and reproducibility research, not biometric identification, surveillance, or inferring sensitive attributes.

⚖️ Licensing

This repository combines metadata created for Déjà Cue with embeddings derived from external research data, so its Hugging Face license field is other. Read DATA_LICENSE.md before reuse or redistribution, and retain the upstream attributions required for the source assets.

📚 Citation

If you use these assets, please cite the official paper record:

@misc{cao2026dejacue,
  title         = {Déjà Cue: Localizing States in Object Histories via Vocabulary-Relative Coordinates},
  author        = {Haofan Cao and Zhichao You and Yunkai Yang and Liang Guo and Jie Wang and Chongshou Li},
  year          = {2026},
  eprint        = {2608.02044},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.02044}
}

Paper: arXiv:2608.02044