A Perceptual Measure for Deep Single Image Camera Calibration

TL;DR

This paper introduces a deep learning approach for single image camera calibration that outperforms existing methods in accuracy and aligns better with human perception, with applications in virtual object insertion and image editing.

Contribution

We propose a deep convolutional neural network for camera calibration from a single image trained on large-scale data, and develop a perceptual measure based on human judgments.

Findings

Our method outperforms existing calibration techniques in L2 error.

The perceptual measure correlates better with human judgments of realism.

Applications include improved virtual object insertion and image compositing.

Abstract

Most current single image camera calibration methods rely on specific image features or user input, and cannot be applied to natural images captured in uncontrolled settings. We propose directly inferring camera calibration parameters from a single image using a deep convolutional neural network. This network is trained using automatically generated samples from a large-scale panorama dataset, and considerably outperforms other methods, including recent deep learning-based approaches, in terms of standard L2 error. However, we argue that in many cases it is more important to consider how humans perceive errors in camera estimation. To this end, we conduct a large-scale human perception study where we ask users to judge the realism of 3D objects composited with and without ground truth camera calibration. Based on this study, we develop a new perceptual measure for camera calibration, and demonstrate that our deep calibration network outperforms other methods on this measure. Finally, we demonstrate the use of our calibration network for a number of applications including virtual object insertion, image retrieval and compositing.