IGEV-Lite Stereo Matching and Pavement-Distress Quantification Code Repository

Description

This dataset contains the complete implementation accompanying the manuscript “Real-time stereo reconstruction and geometric quantification of pavement distress with a variable-baseline platform.” It has been prepared to ensure full reproducibility of all algorithmic, geometric, and quantification results reported in the study. The repository provides a structured and deployment-ready implementation of the proposed IGEV-Lite framework, including the lightweight stereo matching network, the variable-baseline stereo acquisition model, and the plane-referenced quantification pipeline. All modules are arranged to reflect the code dependencies used in the original experiments.

Steps to reproduce

To reproduce the training, calibration, and inference results reported in the manuscript, the following steps can be followed. All scripts and configuration files referenced below are included in this dataset. 1. Environment setup Create a clean Python environment and install the required dependencies using the supplied requirements.txt file. The package lists all Python and PyTorch libraries used in the experiments. A pretrained GhostNetV2 backbone (core/pth/ghostnetv2_100.pth) is provided to ensure consistent initialization. 2. Dataset preparation Public datasets used for training and evaluation (SceneFlow, Sintel, KITTI 2012/2015, and RSRD) should be downloaded from their respective sources. The Datasets/ folder contains YAML configuration files (sceneflow.yaml, sintel.yaml, kitti.yaml, RSRD.yaml) that define dataset roots, training–validation splits, and loading settings. Placing each dataset at the path indicated in the YAML files allows the training pipeline to run without modification. 3. Camera calibration and rectification For reproducing the calibration pipeline described in the paper, run calibrator.py. This script performs intrinsic estimation, distortion correction, and epipolar rectification for a variable-baseline stereo setup. It can also be used to generate rectified stereo pairs for new data collection. The calibration steps mirror those in Section 3 of the manuscript. 4. Training the network The root directory contains the full training implementation for the IGEV-Lite framework. The main model architecture is defined in core/igev_lite.py, supported by the modules in core/utils/ (feature extractor, correlation layers, update units, geometry utilities, and data augmentation functions). Training can be launched by following the procedure described in the README.md: • prepare datasets according to the YAML files • specify training parameters consistent with Section 4 • run the training command to reproduce joint training on synthetic and real data 5. Running inference and evaluation The script testModel.py provides a simple demonstration of running inference on stereo image pairs. It loads the trained network, executes the matching and disparity estimation pipeline, and produces outputs for comparison. The inference configuration mirrors the unified protocol used in the manuscript, including full-resolution inference, update-iteration settings, and ROI-guided processing. 6. Reproducing the quantification pipeline The ROI-based matching, binocular reconstruction, plane fitting, and volume integration described in Sections 3.3 and 4 are implemented directly in the codebase. Running the inference pipeline on rectified stereo pairs will generate maximum depth and integrated volume estimates identical to those reported in the paper. The geometric routines are located in core/utils/geometry.py and related files. 7. Reproducing results in the manuscript

Institutions

• Northeast Forestry University

Categories

Funders

• Science and Technology Project of Heilongjiang Transportation Department

  Grant ID: HJK2023B019

Licence