# Hybrid Deep–Geometric Approach for Efficient Consistency Assessment of Stereo Images

## TL;DR

This paper introduces a method that combines geometry and deep learning to check the consistency of stereo images in real time, without needing calibration data.

## Contribution

A novel hybrid approach that integrates epipolar geometry with Transformer-based detection for real-time stereo consistency assessment.

## Key findings

- HGC-Net reliably detects both severe and mild geometric distortions in stereo image pairs.
- The method achieves high detection rates even at minimal distortion levels and outperforms baseline approaches.
- It operates in real time and supports explainability through confidence and anomaly heatmaps.

## Abstract

What are the main findings?
We propose a self-contained, single-pair stereo consistency check that fuses epipolar geometry with Transformer-based object detection.Our method flags both global camera misalignments and localized semantic or geometric anomalies without external calibration data.

We propose a self-contained, single-pair stereo consistency check that fuses epipolar geometry with Transformer-based object detection.

Our method flags both global camera misalignments and localized semantic or geometric anomalies without external calibration data.

What is the implication of the main finding?
Enables on-the-fly quality assurance of stereo rigs in applications from robotics to 3D cinematography.Lays groundwork for combining semantic scene understanding with classical stereo geometry.

Enables on-the-fly quality assurance of stereo rigs in applications from robotics to 3D cinematography.

Lays groundwork for combining semantic scene understanding with classical stereo geometry.

We present HGC-Net, a hybrid pipeline for assessing geometric consistency between stereo image pairs. Our method integrates classical epipolar geometry with deep learning components to compute an interpretable scalar score A, reflecting the degree of alignment. Unlike traditional techniques, which may overlook subtle miscalibrations, HGC-Net reliably detects both severe and mild geometric distortions, such as sub-degree tilts and pixel-level shifts. We evaluate the method on the Middlebury 2014 stereo dataset, using synthetically distorted variants to simulate misalignments. Experimental results show that our score degrades smoothly with increasing geometric error and achieves high detection rates even at minimal distortion levels, outperforming baseline approaches based on disparity or calibration checks. The method operates in real time (12.5 fps on 1080p input) and does not require access to internal camera parameters, making it suitable for embedded stereo systems and quality monitoring in robotic and AR/VR applications. The approach also supports explainability via confidence maps and anomaly heatmaps, aiding human operators in identifying problematic regions.