---

name: "text-spotting" description: "Guides AdelaiDet BAText/ABCNet text spotting, BezierAlign, text datasets, dictionaries, lexicons, and TextEvaluator workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

text-spotting

Use this sub-skill when a task names BAText, ABCNet, BezierAlign, scene text, text spotting, OCR recognition, lexicons, dictionaries, text evaluation, or Bezier control-point annotations in AdelaiDet.

Use this route for

• Choosing and editing BAText/ABCNet configs.
• Preparing text datasets and Bezier/control-point annotations.
• Understanding BezierAlign, TextEvaluator, custom dictionaries, and lexicons.
• Running text-model demos after setup succeeds.
• Debugging missing recognized strings, evaluator failures, rapidfuzz issues, or text visualization problems.

Do not use this route for

• General installation and CUDA extension failures. Use ../setup-build/SKILL.md.
• Non-text FCOS/BlendMask/CondInst training. Use ../train-eval/SKILL.md.
• General COCO/PIC/LVIS conversion unrelated to text. Use ../data-prep/SKILL.md.
• Checkpoint/ONNX conversion. Use ../export-convert/SKILL.md.

Read first

• references/text-workflows.md for BAText/ABCNet config, training, demo, and evaluation flow.
• references/text-data-and-eval.md for annotation fields, dictionaries, lexicons, and evaluator caveats.
• ../../references/api-reference.md for verified BezierAlign/TextEvaluator surfaces.

Typical workflow

1. Confirm setup with CUDA ops:

   python ../../scripts/check_install.py --cuda-ops
   

2. Choose a configs/BAText/ YAML and inspect MODEL.BATEXT keys.

3. Validate text dataset annotations through data-prep if files are custom.

4. For training/eval, use train-eval launch wrappers and add text-specific overrides.

5. For visualization, use demo-visualize wrappers with the text config/weights.

6. Interpret recognition/evaluation results here, not in generic detection routes.

Decision points

• rapidfuzz.string_metric error: pin rapidfuzz<3 in setup.
• BezierAlign op failure: return to setup-build; text pooling depends on adet._C.
• Boxes draw but text is blank/wrong: check dictionary/lexicon config and annotation transcription fields.
• Dataset lacks Bezier points: route to data-prep; box-only COCO annotations are insufficient for text recognition.
• Evaluation protocol disagreement: identify the target benchmark lexicon/protocol before comparing numbers.