Supplement for "Depth from Combining Defocus and Correspondence Using Light-Field Cameras"
Michael W. Tao, Sunil Hadap, Jitendra Malik, and Ravi Ramamoorthy

ICCV 2013

In this supplemental material, we woud like to showcase our algorithm's robustness both visually and numerically. We are working with noisy and low resolution light-field data. Using both defocus and correspondence depth cues produces better results compared to other methods. The variety of images shows how our algorithm can perform robustly under complex shapes, planes with gradual depth changes, and noise, banding, moire, and repeating patterns.

This supplementary material contains two comparison sections: examples and occlusion boundaries.

Examples
This section shows how our algorithm performs well across different scenarios by utilizing both defocus and correspondence cues.

Yellow Flower

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Purple Plant

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Mural Plane

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Pipe

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Leaf 1

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Plant

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

White Flower

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Plant 2

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Stump

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Leaf 2

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Playground

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Small Flowers

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Purple Flowers

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Pole

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Decorative Flowers

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Shoe

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Guitar

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Guitar and Hat

Input Central View Image*

Defocus Only

Correspondence Only

Our Method

Sun et al. 2010 [1]

Wanner et al. 2012 [2]

Occlusion Boundaries
This section shows how our algorithm performs consistently through multiple examples in detecting occlusion boundaries, yielding higher precision and recall rates.

Decorative Flowers

Input Central View Image*

User-Drawn Occlussion Boundaries

Our Occlusion Boundaries

Sun et al. 2010 Occlusion Boundaries [1]

Wanner et al. 2012 Occlusion Boundaries [2]




Precision: 0.86
Recall: 0.83

Precision: 0.65
Recall: 0.57
Precision: 0.44
Recall: 0.76



Our Depth Estimation

Sun et al. 2010 Depth Estimation [1]

Wanner et al. 2012 Depth Estimation [2]

Purple Plant

Input Central View Image*

User-Drawn Occlussion Boundaries

Our Occlusion Boundaries

Sun et al. 2010 Occlusion Boundaries [1]

Wanner et al. 2012 Occlusion Boundaries [2]




Precision: 0.71
Recall: 0.87

Precision: 0.43
Recall: 0.19
Precision: 0.27
Recall: 0.52



Our Depth Estimation

Sun et al. 2010 Depth Estimation [1]

Wanner et al. 2012 Depth Estimation [2]

Purple Flowers

Input Central View Image*

User-Drawn Occlussion Boundaries

Our Occlusion Boundaries

Sun et al. 2010 Occlusion Boundaries [1]

Wanner et al. 2012 Occlusion Boundaries [2]




Precision: 0.68
Recall: 0.97

Precision: 0.52
Recall: 0.34
Precision: 0.22
Recall: 0.11



Our Depth Estimation

Sun et al. 2010 Depth Estimation [1]

Wanner et al. 2012 Depth Estimation [2]

Playground

Input Central View Image*

User-Drawn Occlussion Boundaries

Our Occlusion Boundaries

Sun et al. 2010 Occlusion Boundaries [1]

Wanner et al. 2012 Occlusion Boundaries [2]




Precision: 0.70
Recall: 0.94

Precision: 0.68
Recall: 0.28
Precision: 0.55
Recall: 0.52



Our Depth Estimation

Sun et al. 2010 Depth Estimation [1]

Wanner et al. 2012 Depth Estimation [2]

Leaf 1

Input Central View Image*

User-Drawn Occlussion Boundaries

Our Occlusion Boundaries

Sun et al. 2010 Occlusion Boundaries [1]

Wanner et al. 2012 Occlusion Boundaries [2]




Precision: 0.72
Recall: 0.77

Precision: 0.74
Recall: 0.52
Precision: 0.63
Recall: 0.30



Our Depth Estimation

Sun et al. 2010 Depth Estimation [1]

Wanner et al. 2012 Depth Estimation [2]

* Input central view images are generated using our light-field processing engine

[1] D. Sun, S. Roth, and M. Black. Secrets of optical flow estimation and their principles. In CVPR, 2010.

[2] S. Wanner and B. Goldluecke. Globally consistent depth labeling of 4D light fields. In CVPR, 2012.

All images are shot with the Lytro Camera in multiple scenarios such as ISO, outdoors/indoors, focal length, and exposure.