Text in everyday settings
Signs, menus, packaging, instructions, and forms photographed in the conditions your product needs to handle.
Bayanat Vision
Image and video data for understanding Arabic in context.
Bayanat Labs (Riyadh) builds Arabic image and video datasets to your scope: signs, packaging, instructions and questions grounded in images or video, labeled by annotators from a team of 50,000+ expert annotators.
اقرأ بالعربيةWhich Arabic image or video data fits your Saudi or UAE project?
Start from what your model has to see, read or point to. Bayanat Vision is built to order, so there is no stock image library; the open release linked here is a synthetic form set for testing a pipeline, not a sample of custom delivery. The first two rows are specific to Saudi Arabia and the UAE; the rest apply in any market.
| Your project | Start with | Check before you commit |
|---|---|---|
| Reading Arabic and English shop signs, menus and street text in Saudi or UAE cities | Custom scene-text scope 10-point Arabic image and video audit | Whether the capture plan covers the lighting, distance, angle, glare and type styles your users meet; whether each text region has a box, a transcription in reading order and a script tag for Arabic, Latin or mixed text; and whether test locations or storefronts stay out of training. |
| Answering shopper questions from bilingual packaging on Saudi or UAE shelves | Custom visual Q&A scope, with an illustrative shelf-photo record | Every answer linked to the region that holds it, a rule for questions the image cannot answer (glare, a hidden label), questions written in Arabic for the task rather than translated from English, and every photo from one store kept in the same split. Source permissions are agreed before collection. |
| Labeling events with timestamps in authorized field-service or in-store video | Custom video event scope | Each event with a start, an end and a stated tolerance; questions about order or change; whether audio or a transcript is in scope; and each recording or session kept whole in one split, never split between neighboring frames. Only authorized recordings are used. |
| Collecting images in places where people may appear | Custom capture plan and permissions | People appear only under the agreed permissions, and the collection and its labels are described without inferring anyone’s identity from appearance. Confirm the hosting location in the scope: in-region is the default, with private-cloud and on-premises options. |
| Licensing a ready-made library of Arabic images or video | Scope a custom collection instead | This family has no stocked library. From any stock supplier, ask for the source and permission behind each image, which images are generated, how many distinct locations or sources there are, and how consent was recorded for people in frame. |
| Mixing generated images with real photos | Custom capture and generation plan | Generated images stay identifiable separately from field collection, so results on real photos can be reported on their own. |
| Pointing to the field a question asks about on a form image | Bayanat Documents Arabic Form Field Extraction (open, synthetic) | Form and invoice fields are a Documents scope with layout and field rules. The open set has 1,152 synthetic form images; its boxes cover label and value areas, not word boundaries, and it has no handwriting or scan degradation, so it is not evidence of accuracy on real forms. |
| Testing a vision-language model on scenes it has not seen | Custom held-out scene splits How to audit an Arabic image or video dataset | Test images from locations, sources or recordings that never appear in training; a pilot sample you draw yourself, stratified by scene and capture condition; and a few dozen items you re-label yourself. Acceptance thresholds stay your decision. |
Is Bayanat Vision the right fit?
A good fit if
- Your model must read Arabic text in images: signs, packaging, menus, instructions or forms.
- You need visual questions and answers tied to visible evidence and coordinates.
- You need event labels and timestamps in authorized video.
Choose something else if
- You want a stock image library. This family is built to order.
- Your task is extracting fields from documents. See Documents.
What you can check before you commission
We separate what is confirmed on this page, what we send on request, and what is agreed for each project.
| Item | Status | Details |
|---|---|---|
| Volume | Per project | Built to order. Image or video count and coverage are agreed with you after a pilot. |
| Label types | Per project | Scene text, bounding boxes, object relationships, answers linked to visible evidence and video timestamps. |
| Capture plan | Per project | Lighting, capture conditions, text styles, layouts and difficulty planned around your task. |
| Real and generated images | Confirmed | Generated images remain identifiable separately from field collection. |
| People in images | Confirmed | We describe the collection without inferring people’s identities from appearance. |
| Example of the label format | On request | Ask for it when you scope your pilot. |
What we can build together
Questions with visible evidence
Arabic descriptions, questions and answers, object relationships, and coordinates that show where an answer comes from.
Events in video
Event labels, timestamps, and questions about sequence or change in authorized recordings.
Relevant variety
Plan lighting, capture conditions, text styles, layouts, and difficulty around the task. Generated images remain identifiable separately from field collection.
How the license covers each use
Training rights do not automatically include display, retrieval or redistribution. Each use is set in the agreement.
Images and video
Source and media permissions are agreed before collection.
Delivered data
Use and delivery rights are set in the agreement.
How we put the data together
Define the use case
Choose a buyer problem, source permissions, and collection plan for a small pilot.
Review the labels
Check text, bounding boxes, relationships, and timestamps. Resolve ambiguous examples before collecting more.
Test new scenes
Separate locations or sources for evaluation where appropriate. Describe the actual collection without inferring people’s identities from appearance.
For example: Reading a product label
A shopper asks which package meets a requirement. The model must read the Arabic labels, locate the relevant information, and answer from what is visible.
Scope your request
Tell us these details. We come back with a fixed-scope pilot plan on your own data: gold-standard QA with adjudication and rework loops and a quality report, usually within two weeks, hosted in-region by default.
- The visual task
- Scenes and text you need covered
- Capture conditions
- Volume
- Timeline
Common questions
Is this a ready-made image library?
This category is offered as custom collection and annotation. We start with a specific use case and pilot, then define the image or video count and coverage with you.
Can the annotations show where an answer appears in an image?
Yes. A custom specification can include scene text, coordinates, object relationships, and answers linked to visible evidence. Video projects can also include event timestamps.
Can you build a Saudi or UAE scene-text dataset?
Yes, as a custom project. We agree the scenes, capture conditions, source permissions and test splits in a fixed-scope pilot with gold-standard QA and a quality report, usually within two weeks, from field capture or media you supply, with capture locations agreed in the scope. Generated images stay identifiable separately from field collection. Confirm the hosting location in the scope: in-region is the default, with private-cloud and on-premises options.
Is there an open Arabic image dataset I can test first?
Not of scenes or product photos: Bayanat Vision is built to order. The open Arabic Form Field Extraction release (1,152 synthetic form images, 9,216 question rows) tests answers linked to field boxes on forms, and the annotation services page shows an illustrative shelf-photo record in the label format. Neither is a sample of custom delivery.