Skip to content

Fragment Segmentation Model

Overview

The AIC-HCMUS Fragment Segmentation Model is designed to identify and segment rock fragments in images. It leverages a YOLOv11-based architecture to detect individual fragments, create precise segmentation masks, and estimate fragment equivalent diameters.

Model Architecture

  • Base Model: YOLOv11m with segmentation capabilities
  • Source: Hosted on Hugging Face Hub (magnusdtd/aic-hcmus-2025-yolo11m-seg)
  • File: yolov11m_finetuned.pt

Key Features

  • Fragment Detection: Precisely identifies individual rock fragments in images
  • Instance Segmentation: Creates accurate pixel masks for each detected fragment
  • Equivalent Diameter Estimation: Calculates the equivalent diameter based on fragment geometry
  • Visualization: Generates overlay images with colored masks for visual inspection

Mathematical Approach

Shape Analysis

For each detected fragment, the model calculates several geometric properties:

  • Circularity: \(C = \frac{4\pi A}{P^2}\)

    • Where \(A\) is the contour area and \(P\) is the perimeter
    • Perfect circles have \(C = 1\)
    • Complex, irregular shapes have \(C \ll 1\)

  • Equivalent Diameter: \(D_{eq} = \sqrt{\frac{4A}{\pi}}\)

    • Diameter of a circle with the same area as the fragment

The equivalent diameter provides a standardized measure of fragment size, allowing for consistent comparison between fragments of varying shapes.

Calibration Detection

For calibration objects (typically red ball), the model analyzes contours using:

\[C = \frac{4\pi A}{P^2} > 0.7\]

Where calibration objects must have high circularity to be considered valid reference objects.

Performance Notes

  • Default execution on CPU
  • Processing time depends on image resolution and fragment count
  • Optimal results with clear, well-separated fragments
  • Recommended image resolution: 512x512 pixels

Limitations

  • Equivalent diameter estimates are based on 2D projections
  • Performance may decrease with crowded or overlapping fragments
  • Best results achieved with good lighting and contrast