Vision & multimodal
Bounding boxes, polygons, segmentation, keypoints, video tracking, OCR and document layout — on data where the text in the image is Japanese and the documents follow Japanese business conventions.
Most vision annotation is language-neutral. The moment there is text in the frame, or a Japanese form on the page, it stops being language-neutral very quickly.
Drawing a box around a car is the same job in any country. Reading the text on a Japanese receipt is not. Japanese documents run vertically as often as horizontally, mix four scripts inside a single line, use full-width Latin characters and digits, carry furigana above the words they gloss, and place a company seal where a Western form expects a signature.
Document AI projects fail on these details rather than on the model. A layout annotation that treats a vertical column as five separate lines, or an OCR transcription that silently converts full-width digits to half-width, produces training data that teaches the model something other than what you intended.
The same is true of on-screen and in-the-wild text: shop signage, product packaging, menus and station boards are dense with Japanese typography that generic annotation guidelines simply do not address.
At a glance
Task types
Geometry, category and text are separate layers, and are reviewed separately.
Bounding boxes, rotated boxes, polygons and pixel-level masks against a class taxonomy agreed before production.
Landmark placement and per-object attributes for pose, product state, condition grading and similar fine-grained work.
Object tracking across frames with stable identities, plus temporal segmentation of actions and events.
Text region detection, reading order, transcription and structural roles on Japanese business documents.
Turning a document into structured values — which text is the total, which date is the issue date, which line items belong together.
Japanese specifics
Four recurring failure points on Japanese image and document data.
Japanese runs top-to-bottom, right-to-left as well as left-to-right, and a single page often mixes both. Tools that assume horizontal lines produce a reading order that is confidently wrong.
How we handle it Reading order is annotated explicitly rather than inferred from geometry, with rules for mixed-orientation pages, ruby text and multi-column layout.
Small kana printed above or beside a kanji word are a pronunciation gloss, not additional content. Transcribed inline, they corrupt the text; discarded blindly, they lose information you may need.
How we handle it Ruby is captured as a linked annotation on its base text, so it can be kept, dropped or used as a reading label downstream without re-annotating.
Ａ１２３ and A123 look almost identical and are different characters. Uncontrolled normalization here is one of the most common silent defects in Japanese OCR datasets.
How we handle it The transcription convention states exactly which characters are normalised and which are preserved, and an automated check enforces it before delivery.
Japanese invoices, receipts and applications have their own structural vocabulary — 御中, 但し書き, company seals, era-based dates, tax-inclusive and tax-exclusive totals sitting side by side.
How we handle it The field schema is built against real examples of your document types, with era dates, seal regions and tax treatment defined as first-class fields rather than free text.
Process
Taxonomy and geometry rules are settled on a real sample before anyone annotates at volume.
We annotate a small representative set to find the ambiguous cases — what counts as one object, what happens at the edge of the frame, when a class is genuinely unclear.
Box tightness, occlusion handling, minimum object size, truncation, group objects and the treatment of unreadable text are all fixed in writing.
Several annotators label the same images independently. Geometry agreement is measured, and where it is low the rule was ambiguous.
Geometry and labels are reviewed as separate passes, because a tight box on the wrong class and a loose box on the right one are different defects.
Annotations ship in your format with the taxonomy version, per-class counts and the QA report for the batch.
Delivery
We export into your training pipeline’s native format rather than making you write a converter.
| Task | Default output | Also available |
|---|---|---|
| Detection and segmentation | COCO JSON | YOLO, Pascal VOC, CVAT XML, masks as PNG |
| Keypoints and attributes | COCO keypoints JSON | CSV, custom JSON schema |
| Video tracking | MOT format | per-frame COCO, CVAT XML |
| OCR and layout | JSON with page geometry | hOCR, ALTO, PAGE XML |
| Field extraction | JSONL, one document per line | CSV, your document-AI schema |
Quality
Vision defects hide well. These checks are the ones that surface them before training does.
FAQ
Common questions on Japanese vision and document AI work.
Yes — handwritten form entries, addresses, names and free-text comments. Handwriting is slower and has a genuinely higher ambiguity rate than print, so we mark uncertain characters rather than guessing, and we agree in advance whether an illegible field is transcribed as best-effort or flagged as unreadable. That decision matters more for your model than the raw accuracy number does.
Yes. We work inside client-provided platforms — CVAT, Label Studio, SageMaker Ground Truth, Labelbox and internal tools among them — or in our own secure environment when you would rather not provision accounts. Which one we use does not change the guideline or the review process.
Yes, and on high-volume detection work it is often the sensible approach. The risk is that correction passes inherit the model’s blind spots, so we measure the correction pass against a from-scratch sample and tell you what it is actually catching. For OCR on Japanese documents we are more cautious, because model errors on full-width characters and vertical text are easy to overlook on screen.
Under NDA and a data processing agreement, in an isolated environment, with access limited to the people on the project. We can also annotate personal information for redaction, or work only on material you have redacted before sending. Retention and deletion terms are agreed in the contract rather than left to policy.
Japanese documents commonly use era years — 令和6年 rather than 2024 — and often mix both on the same page. We treat the era date as its own field with the original string preserved, plus a normalised Gregorian value where you want one, so no information is lost in the conversion and no conversion error is baked into the data.
Image & video
A representative sample of your real documents or footage is enough for us to draft a taxonomy, annotate a pilot set and tell you where the ambiguity will be.
NDA before you share any data. Pilot scoping is free.