A Perceptual Measure for Deep Single Image Camera and Lens Calibration.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Traditional camera calibration is complex, requiring physical targets and multiple images.
• Accurate geometric calibration is crucial for realistic image compositing in digital media.
• Existing single-image methods may not optimize for perceptual realism.

Purpose of the Study:

• To develop a single-image deep learning method for camera calibration.
• To investigate human perception of geometric calibration inaccuracies.
• To create a novel perceptual metric for camera calibration.

Main Methods:

• A deep convolutional neural network was trained on a large-scale panorama dataset.
• A large-scale human perception study was conducted to assess realism with varied calibration parameters.
• A new perceptual measure for camera calibration was developed based on human judgment.

Main Results:

• The deep calibration network achieved competitive accuracy on standard metrics (l2 error).
• The network demonstrated superior performance on the novel perceptual measure.
• Human perception study revealed sensitivities to specific calibration biases.

Conclusions:

• Single-image deep learning offers an efficient alternative to traditional camera calibration.
• Perceptual metrics are essential for optimizing calibration in applications sensitive to visual realism.
• The proposed method enhances applications like virtual object insertion and image compositing.