YLX中

Computer Vision · Geometric Reconstruction · Error Compensation

Learning-based Perspective Distortion Compensation for Industrial Vision Measurement

A Geometry-Constrained Inverse Mapping Framework for Accurate Dimension Recovery

← Back

📐 Projection-to-Geometry Learning for Industrial Vision Measurement

Overview

This project focuses on recovering real-world geometric dimensions from monocular visual observations under perspective distortion.
It aims to bridge the gap between projective geometry and true physical measurements, enabling accurate and robust dimension estimation using low-cost vision sensors.

🔍 Problem Statement

In industrial scenarios, objects captured by cameras suffer from:

  • Perspective distortion (foreshortening effect)
  • View-dependent geometric deformation
  • Nonlinear mapping between image features and real dimensions

Traditional methods based on explicit geometric modeling are often:

  • Sensitive to camera pose
  • Difficult to generalize across setups
  • Limited under real-world noise and depth uncertainty

⚙️ Method

We propose a geometry-constrained learning framework that integrates classical vision geometry with data-driven regression:

1. Planar Reconstruction

  • Extract reference plane using RANSAC
  • Establish a normalized plane coordinate system
  • Reduce 3D problem to structured 2D geometry

2. Perspective Normalization

  • Apply homography-based transformation
  • Align image observations to a canonical top-down view

3. Feature Extraction

  • Use object detection (e.g., YOLO) to obtain:
    • Projected position
    • Projected orientation (yaw)
    • Projected size (length & width)

4. Learning-based Inverse Mapping

We learn a nonlinear mapping:

Projected Geometry→Real Geometry\text{Projected Geometry} \rightarrow \text{Real Geometry}

  • Input: projected geometric features
  • Output: true physical dimensions and pose
  • Model: lightweight regression network (e.g., ResMLP)

🧠 Key Insight

Instead of explicitly modeling the full projection pipeline,
we treat the problem as an inverse geometry learning task, where:

  • Geometry provides structural constraints
  • Learning handles nonlinear residual errors

This significantly improves robustness under real-world conditions.

📊 Results

  • Achieved millimeter-level accuracy (~2 mm error)
  • Strong generalization across varying camera poses
  • Stable performance under noise and partial observation

🚀 Contributions

  • A unified pipeline combining:
    • Planar geometry reconstruction
    • Perspective normalization
    • Learning-based error compensation
  • A data-driven alternative to traditional calibration-heavy methods
  • Demonstration of geometry + learning hybrid paradigm for measurement tasks

📌 Keywords

Computer Vision · Geometric Reconstruction · Error Compensation · Industrial Measurement · Learning-based Inverse Mapping