Computational Rephotography

1
Computational Rephotography

• Soonmin Bae
• Aseem Agarwala
• Frédo Durand

2
What is Rephotography

• The act of repeat photography of the same site
  wikipedia
• Visualize changes over time
• Changes of a city or buildings
• Effects of erosion
• Need to decide
• viewpoint
• season and time of day
• cameras and lenses

3
Newyork changingby Douglas Levere (New York City
photographer)
Custon house statues and New york produce
exchange, Manhattan, 1936
Roadway to the Battery, Manhattan, 1938 and 1997
Native American museum statues and MTA
headquarters, Manhattan, 1997
4
Urban Life though Two Lenses - 1890 and 2002
5
Goals

• Make rephotography easy
• lead photographers to choosing a desired
  viewpoint matching that of a given image
• Visualize the difference between a given image
  and the current scene
• show how to translate and rotate the camera
• Interact with users in a real time

6
Related work

• Depth (or shape) from stereo
• estimate a 3D structure from two images using
  triangulation
• find pairs of corresponding image points and
  measure the difference to infer its depth
• Panoramic mosaic
• increase FOV
• use image-warping with homography (reprojection)
• Visual homing
• lead a robot to a desired positions and
  orientations
• Recover the epipolar geometry relating the
  current image taken by the robot and the target
  image

7
Background

• Homography
• defines a geometric relation between two images
• Epipolar geometry
• find the geometric relation (homography) between
  two images of a single 3D scene
• SIFT (scale-invariant feature transform) Lowe
  04
• extract distinctive features from images
• the features are invariant to image scale,
  rotation, and partially invariant to changing
  viewpoints, change in illumination
• SURF (speeded up robust features)

8
Our Approach

• find and match distinctive features of a given
  image and the scene
• using Computer vision techniques (e.g., SIFT and
  SURF)
• compute the amount of translation
• using epipolar geometry
• interact with users in a real time and lead to a
  desired viewpoint
• using visualization
• resolve rotation
• using image warping with homographies
• or lead users to rotating the camera