Order Accuracy

This project processes a video or RTSP stream, extracts valid order-ID frames, uploads them to MinIO, selects the top frames per order, and runs VLM inference to extract ordered items.

---

📦 What the system does

• Accepts video file uploads or RTSP streams
• Extracts frames using GStreamer + gvapython
• Detects order ID using OCR
• Stores frames in MinIO
• Selects Top-K frames per order using YOLO
• Runs VLM (OpenVINO GenAI) for item & quantity extraction
• Provides a Gradio UI for interaction

---

📁 Project Structure

order-accuracy/
│
├── docker-compose.yaml
│
├── application-service/
│   ├── Dockerfile
│   └── app/
│       ├── main.py               # API + pipeline trigger
│       ├── pipeline_runner.py    # GStreamer launcher
│       ├── frame_pipeline.py     # OCR + frame upload
│       └── requirements.txt
│
├── frame-selector-service/
│   ├── Dockerfile
│   └── app/
│       ├── frame_selector.py     # Selects top frames
│       └── requirements.txt
│
├── gradio-ui/
│   ├── Dockerfile
│   └── gradio_app.py             # Web UI
│
├── config/
│   └── application.yaml
│
├── model/
│   └── Qwen2.5-VL-7B-Instruct-ov-int8/
│
└── storage/
    ├── videos/
    └── uploads/

---

▶️ How to Run

1. Start all services

docker compose up --build

This launches:

• MinIO (frame storage)
• Application Service (GStreamer + OCR + VLM API)
• Frame Selector Service (YOLO ranking)
• Gradio UI

---

Login for MinIO:

minioadmin / minioadmin

---

🎥 How to Use

Upload a Video (UI)

1. Open Gradio UI
2. Upload .mp4 / .avi / .mkv
3. Click Upload & Start

The pipeline starts automatically.

---

RTSP Stream

RTSP example:

rtsp://192.168.1.5:8554/test

API call:

curl -X POST http://localhost:8000/run-video \
  -H "Content-Type: application/json" \
  -d '{"source_type":"rtsp","source":"rtsp://192.168.1.5:8554/test"}'

If localhost is provided in RTSP, the backend safely normalizes it for Docker.

---

🖼 View Frames in MinIO

Extracted Frames

frames/
 └── <order_id>/
      ├── 11.jpg
      ├── 42.jpg
      └── 76.jpg

Selected Frames

selected/
 └── <order_id>/
      ├── rank_1.jpg
      ├── rank_2.jpg
      └── rank_3.jpg

---

🔄 Clean Restart (Recommended)

docker compose down --remove-orphans
docker volume rm order-accuracy_minio_data
docker compose up --build

⚠️ This deletes all stored frames.

---

✅ TL;DR

docker compose up --build
open http://localhost:7860

Upload video or RTSP → frames extracted → top frames selected → VLM results available.