DeepCalib

The implementation of our 2018 CVMP DeepCalib paper. The supplementary material can be found here.

Short description
Requirements
Dataset generation
Training DeepCalib
Camera calibraition
Notes
Citation

Short description

We present a novel fully automatic deep learning-based approach works with a single image of general scenes. Our approach builds upon Inception-v3 architecture: our network automatically estimates the intrinsic parameters of the camera (focal length and distortion parameter) from a general single input image.

Requirements

Python 2.7
Keras 2.1
TensorFlow 1.4
OpenCV 3.2.1

Dataset generation

We provided the code for the whole data generation pipeline. First you have to download sun360 dataset using this download script. There also exists a Google drive link from which you can download manually. Then, you have to choose whether your dataset is going to have continuous or discrete values. We provide the dataset generation code for both in a dataset folder. If you are using this code, please do not forget to cite the paper describing sun360 dataset. If you are unable to download sun360 dataset, here is a link to another dataset of 360 panoramic images.

Training DeepCalib

To train DeepCalib you need to choose which architecture you want to use (refer to the Section 3.3 of our paper). This repo contains all the training scripts for both classification and regression networks as well as all 3 architectures mentioned in the paper. In both regression and classification Seq_Net and Dual_Net folders "dist" and "focal" refer to the netwotks used for distortion parameter and focal length, respectively. All the training codes are available in this folder.

Camera Calibration

To infer distortion parameter and focal length of a given camera we take a short video, extract the frames and run the prediction on all of them. After that, we take the mean or the median of predicted values and use that as a final result. However, in a slight modification you can use them for a single image prediction as well. Below you can see some of the results of image rectification using parameters obtained from single image calibration. In prediction folder we have the codes for all the networks except for SeqNet regression because the weights for this architecture are currently unavailable. We uploaded a simple python script for frame extraction from video sequence.

Notes

Different architectures

For detailed information refer to the Section 4.2 of our paper. In short, SingleNet (a) is the best network for predicting focal length and distortion parameter in terms of accuracy. In addition, since it is a single network contrary to DualNet (b) and Seqnet (c), it is computationally cheaper to use the former.

Weights

The weights for our networks can be found here. We recommend to use SingleNet since we experimentally confirmed it outperforms the other ones. The regression weights for SeqNet are currently unavailable, although you can train your own.

Undistortion

One way to qualitatively assess the accuracy of predicted parameters is to use those to undistort images that were used to predict the parameters. Undistoriton folder contains MATLAB code to undistort multiple images from .txt file. The format of the .txt file is the following: 1st column contains path to the image, 2nd column is focal length, 3rd column is distortion parameter. Each row corresponds to a single image. With a simple modification you can use it on a single image by giving direct path to it and predicted parameters. However, you need to change only undist_from_txt.m file, not the undistSphIm.m.

Citation

@inproceedings{bogdan2018deepcalib,
  title={DeepCalib: a deep learning approach for automatic intrinsic calibration of wide field-of-view cameras},
  author={Bogdan, Oleksandr and Eckstein, Viktor and Rameau, Francois and Bazin, Jean-Charles},
  booktitle={Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production},
  year={2018}
}

@inproceedings{xiao2012recognizing,
  title={Recognizing scene viewpoint using panoramic place representation},
  author={Xiao, Jianxiong and Ehinger, Krista A and Oliva, Aude and Torralba, Antonio},
  booktitle={2012 IEEE Conference on Computer Vision and Pattern Recognition},
  year={2012},
}

mosheliv/DeepCalib