Проект 2: Computer Vision Service

🎯 Цель

Полный backend сервис, обслуживающий YOLO или похожую CV модель в production. Async processing, S3 storage, Docker GPU support — современный CV stack.

Рекомендуемые use cases

**Совет:**License Plate Recognition(узбекский контекст — оригинальный проект) или Crop Disease Detection(PlantVillage готовый dataset).

Architecture

┌─────────────┐
│  Client     │
│  (Web/App)  │
└──────┬──────┘
       │ Upload image/video
       ▼
┌──────────────────────┐
│   FastAPI Backend    │
│  - Auth              │
│  - Validation        │
│  - Routing           │
└────┬───────────┬─────┘
     │           │
     ▼           ▼
┌─────────┐  ┌──────────────┐
│  S3 /   │  │  Celery      │
│  MinIO  │  │  Workers     │
│ (files) │  │              │
└─────────┘  └──────┬───────┘
                    │
                    ▼
            ┌──────────────┐
            │  YOLO Model  │
            │  (GPU/CPU)   │
            └──────┬───────┘
                   │
                   ▼
            ┌──────────────┐
            │  Postgres    │
            │  Results     │
            └──────────────┘

Tech Stack

Required

• **Backend:**FastAPI
• **ML:**YOLOv8 / YOLOv11 (Ultralytics) или HuggingFace
• **Async:**Celery + Redis
• **Storage:**S3 или MinIO
• **Database:**PostgreSQL
• **Container:**Docker (GPU support)

Nice to have

• **Frontend:**Streamlit или React
• **Real-time:**WebSocket
• **OCR:**PaddleOCR
• **Tracking:**Custom (Lightweight DeepSORT)
• **Monitoring:**Prometheus

Features

MVP (1-я неделя)

• Endpoint image upload FastAPI
• YOLO pretrained inference
• Bounding box JSON response
• Возврат annotated image
• Docker (CPU)
• Basic README

V2 (2-я неделя)

• Custom YOLO training (Roboflow или Label Studio)
• S3/MinIO storage (uploaded images, results)
• Celery async processing
• Video upload + frame-by-frame
• Result history (Postgres)
• Tests
• CI/CD

V3 (3-я неделя)

• OCR integration (чтение номеров license plate)
• WebSocket real-time webcam
• GPU Docker image
• Streamlit demo
• Cloud deployment (RunPod / GPU)
• Blog post

API spec

POST /detect/image

curl -X POST -F "file=@photo.jpg" http://api/detect/image

{
    "detection_id": "uuid",
    "detections": [
        {
            "class": "car",
            "confidence": 0.94,
            "bbox": [120, 200, 450, 380],
            "license_plate": "01A123BC"  // OCR result
        }
    ],
    "image_url": "https://s3.../annotated_uuid.jpg",
    "processing_time_ms": 245
}

POST /detect/video (async)

{
    "task_id": "celery_task_uuid",
    "status": "queued",
    "estimated_time_seconds": 120
}

GET /detect/video/{task_id}

{
    "task_id": "uuid",
    "status": "processing",  // queued | processing | completed | failed
    "progress_percent": 45,
    "result_url": null  // когда completed
}

WebSocket /detect/stream

• Browser webcam frame → server
• Server YOLO inference
• Возвращает Bounding boxes JSON (real-time)

POST /annotations (для custom training)

{
    "image_id": "uuid",
    "annotations": [
        {"class": "license_plate", "bbox": [...]},
    ]
}

Project structure

cv-service/
├── README.md
├── docker-compose.yml
├── Dockerfile.cpu
├── Dockerfile.gpu
├── .github/workflows/
├── src/
│   ├── api/
│   │   ├── main.py
│   │   ├── routes/
│   │   │   ├── detect.py
│   │   │   ├── annotations.py
│   │   │   └── ws.py
│   │   └── schemas.py
│   ├── core/
│   │   └── config.py
│   ├── storage/
│   │   └── s3.py                   # MinIO/S3 client
│   ├── ml/
│   │   ├── yolo.py                 # Model wrapper
│   │   ├── ocr.py
│   │   └── tracking.py             # DeepSORT
│   ├── tasks/                      # Celery
│   │   ├── celery_app.py
│   │   └── video_processing.py
│   └── data/
│       └── models.py               # Postgres ORM
├── tests/
├── notebooks/
│   ├── 01_data_exploration.ipynb
│   ├── 02_yolo_training.ipynb     # Roboflow/Colab
│   └── 03_model_evaluation.ipynb
├── data/
│   └── raw/                        # Custom dataset
├── models/
│   └── yolov8_custom.pt
├── frontend/
│   └── streamlit_app.py
└── pyproject.toml

План реализации (3 недели)

Неделя 1 — MVP

• Day 1-2: Dataset collection (фото с телефона или Kaggle)
• Day 3: Annotation в Roboflow (50-200 изображений)
• Day 4: YOLOv8 training (Colab GPU)
• Day 5: FastAPI endpoint + inference
• Day 6: Docker (CPU image)
• Day 7: GitHub + README

Неделя 2 — Async processing

• Day 8: MinIO local setup (Docker)
• Day 9-10: Celery + Redis
• Day 11: Video processing pipeline
• Day 12: Postgres history
• Day 13: Tests
• Day 14: CI/CD

Неделя 3 — Production + Demo

• Day 15: OCR integration (PaddleOCR)
• Day 16: WebSocket real-time
• Day 17: GPU Dockerfile
• Day 18: Streamlit demo
• Day 19: Cloud deployment
• Day 20: Demo video
• Day 21: Blog post

Success metrics

• Detection accuracy (mAP):> 0.85 on custom dataset
• Latency (single image):< 200ms (CPU), < 50ms (GPU)
• **Video processing:**30 fps (CPU), 100+ fps (GPU)
• Concurrent users:> 100 (via Celery)
• OCR accuracy:> 90% on plates

Ресурсы

• Ultralytics YOLO docs — docs.ultralytics.com
• Roboflow Universe — datasets и training
• PaddleOCR docs — multi-language OCR
• MinIO docs — S3-compatible local
• FastAPI WebSocket tutorial

Bonus features

• Multi-model serving — YOLO + OCR + Tracking pipeline
• Custom training UI — upload images → annotate → train (no-code)
• Edge deployment — TensorRT или ONNX runtime
• Mobile app — React Native + image upload
• Real-time tracking — multi-object tracking
• Cost optimization — GPU spot instances

✅ Submission checklist

• GitHub repo
• Custom dataset (100+ images, annotated)
• YOLO custom model fine-tuned
• FastAPI API working
• Async video processing
• OCR integration (если applicable)
• Streamlit demo
• Demo video (web + CLI)
• Blog post
• LinkedIn post

Завершили? Переходите к Проект 3: RAG Chatbot.