Data Labelling App (DLA)

Reference implementation of the Data Labelling App: a pedagogical platform for image classification, object detection, and segmentation that integrates data literacy, supervisor-validated annotation practice, and asynchronous export to COCO and/or YOLO.

Companion manuscript: Educational Annotation Systems for Computer Vision: Automatic Evaluation and Dataset Generation for Medical Applications (docs/article.tex, BRACIS 2026).

---

Overview

DLA couples classroom annotation practice with dataset generation. A supervisor (teacher) registers datasets and exercises; annotators (students) submit labels under:

• Assisted practice — responses compared to a supervisor reference; automatic geometric grading.
• Free practice — labeling without immediate scoring in the same flow.

Validated submissions can be exported as training-ready ZIP archives. The geometric grader (IoU, Hungarian matching, precision/recall/F1) is formalized in the paper; implementation details below supplement the shortened architecture section in the manuscript.

Roles: supervisor ≈ teacher · annotator ≈ student (same individuals, platform vs. classroom vocabulary).

---

Monorepo layout

Path	Role
src/, app.py	Flask API, evaluation domain, Celery workers
frontend/	React web client (annotation UI, dashboard, export) — see frontend/README.md
docker-compose.yml	Full stack: API, frontend, MongoDB, Redis, optional MinIO / observability

Public repository: github.com/ramos-ai/dla-bracis.

---

System architecture

Workflow

1. Supervisor creates datasets and exercises (classification, detection, or segmentation).
2. Annotator submits per-image responses (labelledAnswers for graded practice; unlabelledAnswers for free practice).
3. Server persists submissions and computes supervised scores when configured.
4. Supervisor dashboard aggregates completion, scores, and alerts.
5. Long-running exports are delegated to Celery workers; output is polled by task id.

Reference stack

Layer	Technology
Web client	React, Vite, Nginx (frontend/)
API	Flask, Gunicorn, JWT
Persistence	MongoDB, GridFS or S3/MinIO
Tasks	Redis, Celery
Observability	Prometheus, Grafana (optional profile)
Evaluation	Pure Python (src/domain/evaluation/)

Code layout

Layered backend: HTTP routes → application use cases → domain logic → infrastructure.

src/
├── presentation/      # Blueprints, DTOs, auth, exception mapping
├── application/     # Datasets, exercises, export, reports
├── domain/            # Evaluation (IoU, matching, metrics); exceptions
├── infrastructure/    # MongoDB, Celery, S3/GridFS, security, config
└── shared/            # Logging, utilities

Evaluation module

Component	Module	Role
IoU (bbox + polygon)	iou_calculator.py	Shoelace area; Shapely mask IoU; effective IoU for segmentation
Matching	matching_strategy.py	Hungarian assignment; same-class pairs; cost 1 − IoU
Metrics	metrics.py	Precision, recall, F1 from TP/FP/FN
Orchestration	scoring_engine.py	Classification, detection, segmentation entry points

Constants (ε, ε_seg, A_min) are in constants.py and match the paper.

Export pipeline

ExportConfig (application/datasets/export_config.py): train/val/test splits, format (coco, yolo, both, auto), seed, class filters, optional unlabeled inclusion, resize/JPEG quality. ZIP assembly runs in infrastructure/celery/jobs/export.py.

---

Quick start

Docker (recommended)

cp .env.example .env    # set JWT_SECRET_KEY (required)
docker compose up -d

Service	URL
API	http://localhost:15050
Swagger	http://localhost:15050/api-docs
Frontend	http://localhost:18080

Local API

pip install -r requirements.txt
python app.py

Local frontend (requires running API on port 5000):

cd frontend && npm ci && npm run dev

Observability: docker compose --profile observability up -d (Prometheus :19090, Grafana :13000).

Tests: pytest from the repository root.

---

Configuration

See .env.example. Required: JWT_SECRET_KEY, MongoDB connection. Set S3_STORAGE_ENABLED=true with MinIO/S3 credentials for object storage instead of GridFS-only mode.

Main API groups: /api/auth, /api/dataset, /api/exercises, /api/export, /api/tasks, /api/health.

---

Authors

Pedro da Rosa · Augusto Reich · Felipe Zeiser · Gabriel Ramos
Graduate Program in Applied Computing, Universidade do Vale do Rio dos Sinos (UNISINOS)

---

Funding

Partial support from Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) — grants 313845/2023-9, 443184/2023-2, 445238/2024-0, and 404800/2025-4.

---

Citing

If you use this software in academic work, please cite the companion BRACIS paper (BibTeX to be updated with proceedings metadata):

@inproceedings{darosa2026dla,
  author    = {da Rosa, Pedro and Reich, Augusto and Zeiser, Felipe and Ramos, Gabriel de O.},
  title     = {Educational Annotation Systems for Computer Vision: Automatic Evaluation and Dataset Generation for Medical Applications},
  booktitle = {Brazilian Conference on Intelligent Systems (BRACIS)},
  year      = {2026},
}

---

License

MIT