Structure from Motion

  1. Find keypoints (SIFT)
  2. Find matches (RANSAC)
  3. Align images (Bundle Adjustment)
  4. Dense reconstruction (CMVS)

What does a computer see?

Motion Detector


	  background = video.nextframe()

	  while(video.hasframes()){
	    newframe = video.nextframe()
	    Difference = abs(background - newframe)
	    show(difference)
	  }
	

Scale Invariant Feature Transforms (SIFT)

Edge Detection

Edge Detection

Repeat

Downsample

Repeat

Descriptor

Random Sampling Consensus (RANSAC)

Bundle Adjustment

Bundle Adjustment

Dense Reconstruction

How to run VisualSFM

  • Details and complexity
  • Overlap images
  • Contrast
  • Avoid the sky

Surface texture

Plain Surface

Dotted Surface

Dotted video

Lined Surface

Colored Surface

Binary Surface

Meshlab

Poisson Reconstruction

Texturing



link