Skip to content

Latest commit

 

History

History
137 lines (98 loc) · 4.15 KB

README.md

File metadata and controls

137 lines (98 loc) · 4.15 KB
Raw

Learning-based Camera Calibration

This is the official implementation of the paper: `Learning-based Distortion Correction and Feature Detection for High Precision and Robust Camera Calibration', by Yesheng Zhang and Xu Zhao and Dahong Qian.

Abstract

We propose a learning-based Camera Calibration System (CCS). In this framework, the accuracy and robustness of camera calibration are improved from three aspects: distortion correction, corner detection and parameter estimation. Specifically, the distortion correction is performed by the learning-based method. Accurate feature locations are achieved by the combination of learning-based detection, specially designed refinement and complete post-processing. Moreover, we obtain stable parameter estimation by a RANSAC-based procedure.

MainFig

Paper

Our paper is accepted by RA-L & IROS22. The pdf file can be found here.

BibTex

@ARTICLE{9834080,  
author={Zhang, Yesheng and Zhao, Xu and Qian, Dahong},  
journal={IEEE Robotics and Automation Letters},   
title={Learning-Based Distortion Correction and Feature Detection for High Precision and Robust Camera Calibration},   
year={2022},  
volume={7},  
number={4},  
pages={10470-10477},  
doi={10.1109/LRA.2022.3192610}}

Update log

TODO LIST

  • code released.
  • README completed.
  • Detection weights released. Note this model is trained on basic dataset without augmentation by uneven lighting or distortion. See Issue[#9].

Usage

Requirements

  • CUDA ~= 9.2
  • python ~= 3.6
  • pytorch ~= 1.2.0
  • torchvision ~= 0.4.0
  • python-opencv ~= 3.4.2
  • Numpy ~= 1.16

They can all be installed following command:

    conda create -n CCF python=3.6
    pip install -r requirements.txt
    conda activate CCF

File Folder Configuration

we use the fixed data folder structure for calibration input and output as follows:

data
├── DetectRes # [output] for detection
│   ├── color_img # [output] colored heatmap
│   └── heatmap # [output] heatmap
├── GT # [input and optional] for evaluation
├── SubpixelRes # [output] sub-pixel refinement
├── dist_img # [input] original images
└── img # [output or input] corrected images (output) or images without distortion (input)

See Examples in ./demo_data/*.

Demo data

We provide three sets of calibration data for demo in ./demo_data/.

For simplicity, we directly provide our distortion correction and detection results here.

You can train your own networks for these results using our training scripts as well.

The images provided here are screened out from a larger image set by our RANSAC-based calibration procedure, for the sake of convenience。

Run

First, you need to modify the data path in ./Demo_calib.py , and you can choose one of the three data sets we provided in ./demo_data/.

Then the demo calibration can be run following commands:

    python Demo_calib.py

The results can be seen like:

example

Training

Note: Before the training, you need to prepare the dataset following the Data Generation part and modify the corresponding parameters in settings/settings.py.

We also provide training scripts for our corner detection network and distortion correction network.

Corner Detection: You can use the train_CornerDetect.py, the parameters are as follows:

    python train_CornerDetect.py \
        -e [epoch] \
        -b [batchsize] \
        -l [learning rate]

Distortion Correction: You can use the train_DistCorr.py, the parameters are as follows:

    python train_DistCorr.py \
        -e [epoch] \
        -b [batchsize] \
        -l [learning rate] \
        -o [parameter order]

Data Generation

You can run data_generator.py to generate synthetic dataset with chessboard images and camera parameters.

See dataset/README.md for details.