Deep Single Image Calibration

Thesis Report | Slides | Colab Demo

In this repository, we release our neural network that can estimate from a single image (!) the following camera parameters:

1. Roll (angle of the horizon),
2. Tilt (via offset of horizon from image centre),
3. Focal length (via field of view), and
4. Radial Distortion parameter k1 (via apparent distortion).

One use case of our method is to derive the gravity direction from these outputs and use it as a prior in Sfm or SLAM pipelines.

This project is done as a Master's Semester Thesis at the Computer Vision and Geometry group at ETH Zürich.

Our work builds upon the papers Deep Single Image Camera Calibration With Radial Distortion and DeepCalib: a deep learning approach for automatic intrinsic calibration of wide field-of-view cameras

Quick Start 🚀

We provide a one-liner in quick.py that allows you run our network to calibrate any single image. The colab notebook has a demo that shows you how to quickly get started.

import torch
calibrator = torch.hub.load('AlanSavio25/DeepSingleImageCalibration', 'calibrator')
results = calibrator.calibrate(image_array)
calibrator.visualize(image_array, results)

Under the hood, this performs the required image pre-processing, network inference, and post-processing to derive all the calibration parameters from the network's outputs.

Installation ⚙️

This project requires Python>=3.6 and PyTorch>1.1. The following steps will install the calib package using setup.py and requirements.txt.

git clone git@github.com:AlanSavio25/DeepSingleImageCalibration.git
cd DeepSingleImageCalibration
python -m venv venv
source venv/bin/activate
pip install -e .

Inference

You can run inference using:

python -m calib.calib.run --img_dir images/

This script downloads the weights and saves the network's prediction results + annotated images in results/.

Training

Dataset

We used the SUN360 dataset to obtain 360° panoramas, from which we generated 274,072 images. We use a train-val-test split of 264,072 - 5000 - 2000. To generate these images, first combine all SUN360's images (or any other panorama dataset) into one directory SUN360/total/, then run:

python calib.calib.datasets.image_generation.py

This varies the roll, tilt, yaw, field of view, distortion and aspect ratio to generate multiple images from each panorama.

Then, run the following script to create the train-val-test split:

mkdir -p split
python -m calib.calib.datasets.create_split

Examples of Images in the generated dataset

Distributions of parameters in the dataset

This shows the intervals of the distibutions in the dataset used for training. If a test image has parameters out of this distribution, the network will fail.

Training experiment

Our framework is derived from pixloc, where you will find more information about training. Before starting a training, make sure you edit DATA_PATH and TRAINING_PATH in calib/settings.py. You can modify training configuration in calib/calib/configs/config_train.yaml, if needed. To start a training experiment, run:

python -m calib.calib.train experiment_name --conf calib/calib/configs/config_train.yaml

It creates a new directory experiment_name/ in TRAINING_PATH (set inside settings.py) and saves the config, model checkpoints, logs of stdout, and Tensorboard summaries.

--overfit flag loops the training and validation sets on a single batch (useful to test losses and metrics).

Monitoring the training: Launch a Tensorboard session with tensorboard --logdir=path/to/TRAINING_PATH to visualize losses and metrics, and compare them across experiments.

Interpreting the neural network

We also add a notebook (visualize_layercam.ipynb) to visualize the gradients of the network. Before using this, run the following:

cd class_activation_maps
python layercam.py -c path/to/config_train.yaml --head roll -e exp12_aspectratio_5heads

We adapt code from here. Notice that the network focuses on straight lines and horizons.