How joker-qp/ando-sign3d works
This repository contains an ASL fingerspelling research workflow: curate hand-image data, prepare normalized landmarks, train a 26-letter classifier, evaluate it, and export model artifacts. A live-camera inference module is present, though its implementation was not sampled, so its runtime details and wiring are limited to documented responsibilities. The sampled sources support the training pipeline’s use of curation, normalization, model, and academic-support modules; the tree and model documentation identify the remaining product capabilities.
Main entry points: train.py Training pipeline · m01_asl_letters Model and reports · io_manifest_reader.py Curation and splits · test_cam.py Camera inference
Components 16
Data preparation
- Curation and splits
src/common/io_manifest_reader.py - Curation rules
configs/preprocessing_rules.yaml - Curated manifests
- Landmark preparation
src/models/m01_asl_letters/train.py - Pose normalization
src/common/math_normalizer.py
Model workflow
- Training pipeline
src/models/m01_asl_letters/train.py - Model settings
configs/models/m01_asl_letters.yaml - ASL classifier
src/models/m01_asl_letters/model.py
Research outputs
- Academic evaluation
src/common/academic_evaluator.py - Experiment telemetry
src/common/academic_logger.py - Model card generation
src/common/report_generator.py - Model and reports
artifacts/m01_asl_letters - Dataset QA report
src/tools/generate_dataset_qa_report.py
Live inference
- Camera inference
src/models/m01_asl_letters/test_cam.py
Other components
- ASL images
- Camera user
Connections 21
- ASL images → Curation and splits: provides images
- Curation rules → Curation and splits: configures
- Curation and splits → Curated manifests: writes manifests
- Training pipeline → Curation and splits: runs if missing
- Curated manifests → Training pipeline: loads data
- Training pipeline → Landmark preparation: extracts landmarks
- Landmark preparation → Pose normalization: normalizes poses
- Model settings → Training pipeline: configures
- Training pipeline → ASL classifier: trains and exports
- Training pipeline → Experiment telemetry: logs experiment
- Training pipeline → Academic evaluation: evaluates model
- Training pipeline → Model card generation: generates card
- Training pipeline → Model and reports: writes outputs
- Academic evaluation → Model and reports: writes evaluation
- Experiment telemetry → Model and reports: writes metrics
- Model card generation → Model and reports: reads run outputs
- Curated manifests → Dataset QA report: reads manifests
- Camera user → Camera inference: provides camera input
- Pose normalization → Camera inference: normalizes poses
- ASL classifier → Camera inference: classifies landmarks
- Model and reports → Camera inference: provides exported model