🍲 Food & Vessel Detection Pipeline

A modular computer vision pipeline designed for reference-free volume estimation. It detects vessels, calculates real-world dimensions using metric depth maps, and identifies liquid levels even in transparent containers.

---

🚀 Key Features

1. Metric Depth Estimation (Reference-Free)

The pipeline uses Depth-Anything-V2-Metric to predict absolute distance in meters for every pixel in the scene. Unlike traditional systems, it does not require a known reference object (like a coin) or hardcoded vessel sizes.

2. Camera Intrinsics (EXIF Parsing)

Accuracy in size estimation depends on the camera lens. The system automatically:

• Reads EXIF metadata (Focal Length, 35mm Equivalent) from your photos to calibrate the "vision" scale.
• Falls back to a standard mobile FOV (65°) if metadata is missing.
• Allows manual calibration via the --fov argument.

3. Hybrid Liquid Level Detection

For transparent containers where standard segmentation fails, a specialized processor analyzes depth gradients and meniscus edge profiles to find the water/drink line.

4. Geometric Volume Modeling

Volume is calculated in milliliters (ml) by mapping 2D masks to 3D approximations:

• Cups/Glasses: Modeled as vertical cylinders ($V = \pi r^2 h$).
• Bowls: Modeled as hemi-ellipsoids ($V = \frac{2}{3} \pi r^2 h$).

---

🛠 Setup & Usage

Docker (Recommended)

This project uses uv for ultra-fast dependency management inside the container.

1. Start the persistent container:

   docker-compose up -d --build

2. Run detection:

   docker exec food-volume-detection python main.py --path data/glass.png

Command Line Arguments:

• --path: (Required) Path to the input image.
• --fov: (Optional) Manual Field of View override. Use 30-40 for zoomed photos or 65-75 for standard wide shots.
• --distance: (Optional) Manual distance in cm (e.g., --distance 30).
• --device: (Optional) cpu, cuda, or mps.

---

📂 Project Structure

• src/:
  • detection/: YOLOv10 (vessel) and YOLOv8-seg (surface) inference.
  • depth/: Metric depth prediction using Hugging Face Transformers.
  • processing/: Hybrid liquid level detection logic.
  • metrics/: EXIF parsing and geometric volume calculation.
  • utils/: Structured results saving and visualization.
• results/: Persistent logs organized by image name (e.g., results/image_name/result.jpg).
• main.py: Clean entry point for orchestration.

---

💡 Best Practices for Accuracy

• Resolution: Use high-resolution (1080p+) original photos from your phone.
• Metadata: Do not strip EXIF data (e.g., via messaging apps) before inputting to the pipeline.
• Zoom: If your photo is cropped or zoomed, use the --fov override (e.g., 30) to compensate for the digital zoom.