ML2/navocr-testsets

NavOCR Test Sets

Test sets for navigation-relevant text detection, introduced in NavOCR: A Dataset Generator for Navigation-Relevant Text Detection in Mobile Robots (NeurIPS 2026 Evaluations & Datasets Track). Code: https://github.com/kc-ml2/NavOCR

"Navigation-relevant" text is text a person or robot would use to say where they are, mostly shop and place names. Menus, prices, slogans and opening hours are not navigation-relevant.

Benchmark	Images	Annotation	Purpose
Detection Benchmark	1,554	Bounding boxes, COCO format (1,682 boxes)	Locating navigation-relevant text in an image
Text Selection Benchmark	304 (subset of the Detection Benchmark)	Text strings per image	Choosing the navigation-relevant strings among all text read from an image

Files

detection_benchmark/
  images/                 1,554 JPEG images
  instances_test.json     COCO annotations, one category: navigation-oriented-text
text_selection_benchmark/
  images/                 304 JPEG images, identical to their Detection Benchmark copies
  text_string_label.txt   one line per image
croissant.json            Croissant metadata (core + RAI)

Detection Benchmark: instances_test.json

Standard COCO detection format. bbox is [x, y, width, height] in pixels of the released image. Each box marks one text instance.

Text Selection Benchmark: text_string_label.txt

One line per image: the file name, then three brace fields.

20251215_135837.jpg, {H&M, CHARLES & KEITH, McDonald's}, {H&M, 찰스앤키스, 맥도날드}, {에이치앤엠, Charles & Keith, McDonald's}

1. The navigation-relevant strings exactly as written on the signs. This is the label to score against.
2. The names of the places in the image as listed in open map resources, such as OpenStreetMap.
3. Translations of those names (Korean to English or English to Korean). Map resources list place names in the local language, while signs are often written in English, or the other way round, so a sign and its map entry can differ only by language.

Annotation guideline

Detection Benchmark boxes follow these rules:

• One text instance per place. When the same shop shows several signs, only one is annotated.
• Different places get separate boxes.
• Text more than about 20 m away is not annotated.
• Text on surfaces more than 50° off the viewing direction is not annotated.

Two of the authors annotated the images. Each image was annotated by one person, who drew the boxes by hand in a custom tool, with no OCR pre-annotation. In the Text Selection Benchmark, the on-sign strings and the map names were written by hand; the translations in field 3 were generated by an AI model.

Collection

The authors photographed storefronts and street scenes with smartphones, from a pedestrian's point of view, between October 2025 and October 2026. Most images are from South Korea, mainly Seoul and Daejeon; a few show signs abroad. Most scenes are shopping malls and street-level shops. Phones: iPhone 14 Pro, iPhone 15 Pro, iPhone 17, Galaxy S24+ and Galaxy S26.

Every image was capped at 1080 px on the long side before annotation (Lanczos, JPEG quality 95), so the released resolution is the annotated resolution. Most images are 1080×810 or 810×1080; 191 are 360×480 and 13 are 1024×768.

Privacy

• Faces are blurred. Detector proposals (CenterFace via deface, and SCRFD-10G) were reviewed image by image for all 1,554 images. Missed people were added, including people seen from behind or from the side, and false positives were removed. 2,050 faces in 711 images are blurred. No blur overlaps an annotated text box.
• Faces on advertisements, posters and mannequins may remain visible.
• Licence plates are not masked.
• All metadata is removed: EXIF (including GPS, which 204 originals carried), XMP, and the second image embedded in MPO files. Images without faces keep their original pixels.

If you find an identifiable person or other content that should be removed, contact chlee@kc-ml2.com and it will be removed in a new revision.

Intended use

Use these sets to evaluate:

• detectors and OCR pipelines that find navigation-relevant text (Detection Benchmark, e.g. COCO AP);
• methods that pick place names out of recognised text (Text Selection Benchmark, e.g. F1 over strings).

Do not use them to identify, track or recognise people, for surveillance, or for any commercial purpose.

Limitations and biases

• Mostly South Korea (Seoul, Daejeon), mostly commercial areas and indoor malls. Text is mostly Korean and English.
• Taken with five phone models.
• Many places are chain brands.
• 191 images are low resolution (360×480).
• The annotations follow the navigation-relevance definition above, not all visible text.

Versions

• v1.0.1: removed 13 duplicate boxes (1,695 → 1,682). Results in the NavOCR paper use v1.0.

License

CC BY-NC 4.0. Non-commercial use only, with attribution.