BSTRB contains photographs and video frames of everyday objects recorded by study participants in their homes. Access is granted for non-commercial research to people affiliated with a research institution. Requests are reviewed manually.

By requesting access you agree to: (1) use BSTRB only for non-commercial research; (2) not attempt to identify, contact or locate any participant or any person, address or account that may appear in the data, and not link the data with other sources for that purpose; (3) not redistribute the images, in whole or in part, outside your research group (publishing derived statistics, model weights trained on BSTRB, and up to a few example images in papers is allowed); (4) report any personal information you find in the data to the authors so it can be removed; (5) delete the data on request. Violations end your access.

BSTRB: Blurry Scene Text Recognition Benchmark

BSTRB measures how well vision-language models read degraded text: small, blurred, far-away text on everyday objects, as captured by phones and smart glasses. Each object was filmed while the camera moved away from it, so the same text appears at a sequence of shrinking sizes; a separate set of photographs covers close-up and distant shots.

Paper · Code · Project page · Hawkeye-4B · Hawkeye-8B

At a glance

split	images	video frames (videos)	photos	labels
`train`	7,399	6,704 (1,128)	695	GPT-5 transcriptions
`validation`	369	333 (56)	36	human-verified
`test`	1,566	1,325 (223)	241	human-verified

No video appears in more than one split, and no image is repeated.
The public splits omit a few units removed in release review because they showed personal information (account screens, network credentials, stored-value cards, personal photographs, a private event program): 30 train frames (5 videos) and 5 train photos, and one test video (6 frames). The paper's test numbers include that video (1,331 frames); on the public test split (1,325 frames) every model's video BoW F1 moves by at most 0.001 and the ranking is unchanged.
validation and test together are the human-verified ("golden") set. The authors selected their checkpoints on validation; report results on test.
All frames of one video share its transcript, so frames are not independent: compute confidence intervals by resampling videos (video_id), not frames.

from datasets import load_dataset
ds = load_dataset("byron123t/BSTRB")          # after accepting the agreement + huggingface-cli login
ex = ds["test"][0]
ex["image"], ex["prompt"], ex["transcription"]

Fields

field	description
`image`	the image (JPEG, a few PNG)
`id`	`bstrb-v-NNNNN-fFFFFF-<position>` for a video frame, `bstrb-p-NNNNN` for a photo
`split`	`train`, `val` or `test`
`kind`	`video_frame` or `photo`
`video_id`	`bstrb-v-NNNNN` (frames only): the unit for clustering
`frame_position`	where in the video the frame was sampled: `first`, `start`, `early`, `middle`, `late`, `end` (the camera moves away, so text shrinks along this sequence)
`frame_index`	the same as 0–5
`text_height_px`	median text height in the frame, in pixels (frames only). Measured by text detection on the earlier frames and propagated to the later, smaller ones (see the paper's appendix)
`resolution`	`high` or `low` capture of a photo pair (photos only)
`prompt`	the instruction given with the image (several phrasings of "transcribe the text")
`prompt_category`	`StudyPromptCropped`, `FormatConstrained` (asks to keep line structure) or `PhotoStudyOCR`
`transcription`	the reference text, line breaks preserved
`label_source`	`human_verified` (val/test) or `gpt5_auto` (train)
`transcription_auto`	the original GPT-5 transcription (identical to `transcription` for train)
`exif_orientation`	the camera's original EXIF orientation flag; see below

How it was made

Collection. 244 adults in the United States were recruited through Prolific (IRB-exempt study, informed consent, paid ~US$9.93/hour). Each filmed objects carrying at least ten words of text (packaging, books, signs, screens) while walking backwards for about eight seconds; pilot participants instead took a close-up and a distant photograph of each object. Participants were told not to film anything with personal information (mail, bills, prescription labels, faces).

Frames. From each video, 5–6 frames are sampled along the walk-back (frame_position) and the text region is cropped; as the camera retreats, the crop covers fewer pixels, which is what makes later frames hard.

Labels. Every image was first transcribed by GPT-5. For the golden set, one annotator reviewed each video's (or photo's) transcript against its images and corrected it: 20% of video transcripts and 56% of photo transcripts changed (automatic photo labels were often incomplete). Two video units the annotator could not verify were dropped.

Training labels caveat. train keeps the GPT-5 labels. For photos these are known to be systematically incomplete (median 102 vs 342 characters after human review on the golden photos), which is why the Hawkeye models were trained on video frames only.

Privacy

Units with third-party personal information found by an audit of the transcripts (a school form, a prescription label and its medication carton, a business card) were removed.
Released ids are neutral; the original filenames (which carried device names and study ids) are not released.
All image metadata (EXIF including GPS, XMP, IPTC, comments) was removed. The compressed image data was copied byte for byte, so pixels are exactly what models were trained and evaluated on.
Every image was screened with a face detector (SCRFD); the 126 videos/photos with a detected face were reviewed by hand and those showing real people or a private event were removed.
If you find personal information, contact the authors (bjaytang@umich.edu) and it will be removed.

Orientation

Images are stored in the camera's sensor orientation, which is what the models saw (neither training nor evaluation applied the EXIF rotation). 670 of the 977 photos had a rotation flag; it is kept in exif_orientation (6 = rotate 90° clockwise to view, 3 = 180°, 8 = 90° counter-clockwise). Video frames are upright.

Evaluation protocol (as in the paper)

Metrics are computed per image after lowercasing, removing punctuation and collapsing whitespace: bag-of-words F1 (headline), word inclusion rate, WER and aligned F1; see hawkeye/metrics.py. Comparisons between models are paired on the same items with a bootstrap that resamples videos (photos are their own unit), Holm-corrected across comparisons.

BoW F1 on the public test video frames (1,325): Hawkeye 8B 0.743, Hawkeye 4B 0.727, Qwen3-VL-235B 0.725, Kimi K2.6 0.722, GPT-6 Astra 0.708, base Qwen3-VL-8B 0.674, base Qwen3-VL-4B 0.626 (the paper's figures on 1,331 frames are 0.001 higher).

Known limitations

Mostly English text on US consumer products; a few non-Latin scripts.
Faces that remain are printed ones (packaging, book and album covers, posters). The face detector can miss small or blurred faces, and other incidental content (screens, house numbers) was screened only through the transcripts. Other incidental content (screens, house numbers) was not screened automatically.
One annotator verified the golden labels.

License and citation

CC BY-NC 4.0, plus the access agreement above. The images were contributed by study participants under the consent terms of the study; do not use them to train systems that identify people.

@article{tang2026hawkeye,
  title   = {Hawkeye: Reading Illegible Text with Vision Language Models},
  author  = {Tang, Brian and Shin, Kang G.},
  journal = {arXiv preprint arXiv:2610.41254},
  year    = {2026}
}

Downloads last month: 14

Total file size:

4.56 GB

Models trained or fine-tuned on byron123t/BSTRB

Paper for byron123t/BSTRB

Hawkeye: Reading Illegible Text with Vision Language Models

Paper • 2610.41254 • Published 5 days ago