WearerTextBench

WearerText: Benchmarking Wearer-Centered Scene Text Understanding in AI-Glasses Videos

Project & code · Project page

Release scope

This repository contains the test split only: 107 videos and 1,391 QA pairs, with 107 examples for each of 13 tasks. Each video is stored once and referenced by 13 metadata rows. Videos occupy 2,471,387,363 bytes.

The original files use MPEG-4 Part 2 video (mp4v), which is not supported by all browsers. A separate preview/ folder contains H.264 720p display copies for the project homepage. These are not evaluation inputs and are excluded from the configured test split. Original videos are retained byte-for-byte under test/. Download originals to reproduce experiments; do not substitute previews.

The manuscript describes a full benchmark of 1,101 videos and 15,336 QA pairs. Its training split (994 videos, 13,945 QA pairs) is not included in this release. Do not infer that those training data are downloadable here.

Loading

Requires Python 3.9+ and datasets>=4,<5.

from datasets import load_dataset, Video

ds = load_dataset("AI4Reading/WearerTextBench", split="test")
# No video decoder needed when only reading paths and annotations.
ds = ds.cast_column("video", Video(decode=False))
print(ds[0])

For local validation before upload:

ds = load_dataset("videofolder", data_dir="/absolute/path/to/huggingface",
                  split="test")
ds = ds.cast_column("video", Video(decode=False))
assert len(ds) == 1391

Decoding frames through datasets.Video additionally requires a compatible TorchCodec/PyTorch/FFmpeg installation. Alternatively open the MP4 files using your preferred video reader. The plain seven-field annotation list is also available as annotations/test.json.

Schema

One example represents one question, not one distinct video.

Field	Meaning
video	Hugging Face Video feature; shared MP4 reference
video_path	Original MP4 basename
question_id	Unique corrected ID: {level}_{task_id}_{original_suffix}
task_id	One of the 13 final task IDs below
task_name	Full task name and acronym
level	Integer 1, 2 or 3
question	Original-language wearer-centered question
answer	Verified reference answer

test/metadata.jsonl contains these annotation fields plus file_name, the VideoFolder join key; the loader converts it to the video feature. No model predictions or judge outputs are bundled as ground truth.

Task hierarchy

Level	ID	Task
Text Perception	L1.1	Dynamic Text Recognition (DTR)
Text Perception	L1.2	Dynamic Text Localization (DTL)
Contextual Understanding	L2.1	Query-Guided Text Extraction (QTE)
Contextual Understanding	L2.2	Confirmation (CF)
Contextual Understanding	L2.3	Cross-Lingual Understanding (CLU)
Contextual Understanding	L2.4	Situated Semantic Interpretation (SSI)
Contextual Understanding	L2.5	Wearer-Centered Situational Inference (WSI)
Wearer-Grounded Reasoning	L3.1	History-Conditioned Intent Inference (HII)
Wearer-Grounded Reasoning	L3.2	Action-Trajectory Intent Prediction (AIP)
Wearer-Grounded Reasoning	L3.3	Structured Text Synthesis (STS)
Wearer-Grounded Reasoning	L3.4	Spatio-Temporal Navigation Reasoning (STN)
Wearer-Grounded Reasoning	L3.5	Goal-Oriented Comparison (GOC)
Wearer-Grounded Reasoning	L3.6	Viewpoint Trajectory Description (VTD)

Legacy compatibility: old L2.5 → current L3.3; old L3.3 → current L2.5. Update the level prefix too, preserving the entire original suffix. The data here are already corrected. Never apply that swap twice.

Collection and verification

The manuscript describes real-world RayNeo AI-glasses recordings in shopping, transportation and tourist scenarios. It uses layout-aware reference text, hybrid question generation, and multi-agent plus human verification (MAH-V). The eight recording volunteers were members of the research team and provided informed consent. The manuscript reports anonymization of third-party faces.

These statements describe the authors' documented process, not an independent privacy audit of every frame in this upload package. The release owner must confirm that the packaged files are the approved anonymized release copies.

Evaluation

Use the project's VerEval implementation with ground-truth annotations and model predictions. Final score: 0.9 × Consistency + 0.1 × Logic. Fact is an intermediate diagnostic, not a third equally weighted component. Report judge model/version, evidence mode, video-frame sampling, prompt version and full test coverage. The paper's main table uses evidence-based mode (question, prediction, verified ground truth; no video frames) with a local Qwen3-VL-32B judge, temperature 0, a 1,024-token cap and a deterministic seed. Video-aware judging is an optional alternative and must be reported separately.

Intended use and limitations

Academic, non-commercial research, with attribution and responsible handling; see use policy. No de-anonymization or identification attempts. Public-scene and language coverage are bounded by the authors' recording locations and device. Camera-view trajectory is not eye gaze. Short-clip offline QA does not establish reliable long-horizon navigation or deployment safety.

Integrity and provenance

manifest.json records the bytes and SHA-256 of every video and metadata file. preview-manifest.json separately records browser previews and source hashes. release-summary.json records exact released counts. The package is derived from the supplied corrected test annotations, without rewriting questions or answers. The code repository retains the source task definitions and results table for provenance.

The project-manuscript citation is available in the code repository at data/citation.bib. No archival venue identifier or DOI is asserted by this release card.