Accept the TerraSense Analytics Data Release Terms to access RAVEN-Bench
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
The RAVEN-Bench videos and annotations are provided by TerraSense Analytics Ltd. under the TerraSense Analytics Data Release Terms (see DATA_LICENSE.md): research and non-commercial use only, attribution required, no re-identification of individuals, no surveillance, tracking or profiling of specific individuals, and downstream recipients must be bound by the same terms. By requesting access you agree to these terms.
Log in or Sign Up to review the conditions and access this dataset content.
RAVEN-Bench
A Paired EO–IR Video QA Benchmark for Aerial Multimodal Understanding (NeurIPS 2026, Track on Evaluations and Datasets) · Code · Citation
RAVEN-Bench contains 100 temporally aligned EO–IR video pairs (200 MP4 streams) and 1,200 human-authored four-way multiple-choice questions, twelve per pair. Each pair has an eight-question consistency group (C1) and a four-question coherence group (H1). Models are evaluated under EO-only, IR-only and paired EO+IR input.
Version: 1.0.0
Access
Access is gated. Sign in, state your affiliation and intended research use, and accept the TerraSense Analytics Data Release Terms. Requests are approved by the maintainers. After approval, the dataset files are available from this repository.
License and required attribution
The videos and annotations are provided by TerraSense Analytics Ltd. under the TerraSense Analytics Data Release Terms: research and non-commercial use only. Re-identifying individuals, surveillance, tracking or profiling of specific individuals, and the other uses listed in section 4 are prohibited. Anyone who redistributes the dataset or a modified version must pass the same terms on to recipients and identify the modifications. Results such as model outputs and trained weights are not restricted. Evaluation, example and tool code is licensed under the Apache License 2.0; see LICENSES.md.
Any publication, presentation or distribution that uses the dataset must include:
This work uses data provided by Terra Sense Analytics Ltd. (www.terrasense.ca).
Composition
| Item | Count |
|---|---|
| EO–IR video pairs | 100 |
| EO / IR streams | 100 / 100 |
| Questions (12 per pair, 4 options each) | 1,200 |
| L1 / L2 / L3 questions | 386 / 560 / 254 |
| EO / IR / EO+IR required evidence | 656 / 229 / 315 |
| Consistency / coherence groups | 100 (8 items) / 100 (4 items) |
| Total stream duration | 17,939.0 s (mean 89.7 s per stream) |
| Scenario families | 59 ground transport and machinery · 27 maritime · 11 aircraft · 3 people |
| Video bytes | about 158 GB |
The default Hub view has one row per pair with two video columns. Its test label names the evaluation
collection; there is no train or validation split.
Files
videos/metadata.jsonl # 100 paired rows with EO and IR references
videos/eo/*.mp4 # TerraSense_eo_1.mp4 ... TerraSense_eo_100.mp4
videos/ir/*.mp4 # TerraSense_ir_1.mp4 ... TerraSense_ir_100.mp4
data/annotations/*.json # one annotation file per pair (source option order and answers)
data/questions.jsonl # 1,200 evaluation rows with the benchmark option permutation
data/model_inputs.jsonl # the same questions/options without answers or hints
data/pairs.jsonl # pairing index
data/captions.jsonl # event descriptions (not model input)
data/metadata/pairs.csv # pair index used by the evaluation code
data/questions.csv # flat question view used by Croissant
croissant.json # Croissant metadata with Responsible AI fields
release_manifest.json # membership and file hashes
checksums.sha256 # content checksums, including the 200 MP4 files
See docs/DATA_FORMAT.md for field definitions.
Evaluation
The included evaluation code reads this dataset and exports requests for all three input settings. Its configured input policy uses the full videos at 1 frame per second, no default frame cap, and a 262,144-pixel area budget per frame. See the input protocol and report the actual policy used with results.
Send only the question, the options and the selected video input to a model. Do not send captions,
answers, option roles, capability labels, question types, modality requirements or evidence intervals.
Annotation files use the source option order; data/questions.jsonl and data/model_inputs.jsonl use
the benchmark's deterministic permutation (eo_ir_option_shuffle_v1), so score against the
correct_answer in the question table.
To inspect video references without decoding:
from datasets import Video, load_dataset
ds = load_dataset("TerraSenseAnalytics/RAVEN-Bench", split="test", streaming=True)
ds = ds.cast_column("eo", Video(decode=False))
ds = ds.cast_column("ir", Video(decode=False))
pair = next(iter(ds))
print(pair["sample_id"], pair["eo"], pair["ir"])
python examples/download_sample.py fetches four example pairs (about 415 MiB) with their annotations.
Responsible use
RAVEN-Bench is intended for research evaluation of EO–IR video reasoning, not for operational deployment. The release contains no boxes, masks, tracks, trajectories, geolocation, flight paths or operational metadata. See docs/RESPONSIBLE_AI.md for provenance, privacy screening, limitations and residual risks.
Issues
Report data problems through the GitHub issue tracker or the Community tab of this dataset. For access or licensing questions, contact the corresponding authors, Zheng Liu (zheng.liu@ubc.ca) and Mohsen Zardadi (mohsen.zardadi@terrasense.ca).
Included evaluation code
The code directory contains the portable evaluation engine, model runners and configs. Install with python3 -m pip install -e ./code; default preparation is offline.
- Downloads last month
- 232