Magnum stock detector model

This is a tool to improve and test computer vision model for magnum ice cream company

🛠 Project Architecture

This repository is structured into three main component directories:

.
├── 📁 01_yolo_detection_training/     # YOLO Object Detection Training Pipeline
├── 📁 02_yolo_segmentation_training/  # YOLO Instance Segmentation Training Pipeline
└── 📁 03_MVP/           # Tool to test and improve model

---

📂 Module Descriptions

1. 01_yolo_detection_training/

Contains the scripts, configuration files, and dataset processing utilities required to train and evaluate the YOLO Object Detection Model.

• Purpose: Identifies and locates individual product SKUs (bounding boxes) within commercial freezers.

2. 02_yolo_segmentation_training/

Contains the training pipeline for the YOLO Instance Segmentation Model.

• Purpose: Generates precise pixel-level masks for structural elements inside the freezer (such as grid dividers, baskets, lanes, and shelves). These masks define spatial boundaries used for depth calibration.

3. 03_MVP/

The core client-facing application and testing suite built using FastAPI (Backend) and Flet (Frontend UI).

---

🚀 Key Features

• 3D Depth-Based Inventory Estimation: Combines 2D object detection with 3D depth maps to estimate hidden or stacked items behind visible stock.
• Granular SKU Analytics: Generates both high-level aggregated inventory totals and individual detection metrics.
• Interactive Dashboard: Dark-themed UI for uploading imagery, previewing bounding boxes, viewing segmentation polygons, and auditing depth maps in real time.

---

💻 Tech Stack

• Computer Vision & ML: Ultralytics YOLO (Detection & Segmentation), Hugging Face Transformers (Depth Anything V2), OpenCV, PyTorch.
• Data Processing: SciPy (cKDTree for spatial $k$-NN search), NumPy, Pillow.
• Backend: FastAPI, Uvicorn.
• Frontend UI: Flet (Flutter for Python).