Depth from Stereo Voicu Popescu Matt Waibel Comp 290-075 5-01-2000.

Depth from Stereo

Voicu Popescu

Matt WaibelComp 290-075

5-01-2000

Overview

• Image Acquisition

• Camera Calibration

• Epipolar Geometry

• Correlation

• Stereopsis

• Ideas for Future Work

Image acquisition: Timbuktu

Camera Calibration

Calibration: image points

Calibration: 3D points

Camera Calibration

Cal_L.0001.log Cal_R.0001.log

Camera Calibration

• 60 deg horizontal field of view

• 35 mm film

• 8 inches baseline

• 4096 x 3112 pixels

Epipolar Geometry

][81,31maxmin,1

maxmin,2

maxmin,/7207991maxmin,

2

][2.203],[35],[3112],[4096

[deg]60

mmeez

vv

zuu

mmbmmwfilmpixhrespixwres

fovh

Timbuktu

Timbuktu

Helicopter

Eurotown

Correlation (Template Matching)

• Choosing Template size (from left image)– dependant on feature size

Correlation (cont.)

• Mean squared error of each channel

• Determining the “best” match from the correlation

Image Preprocessing

• Resizing images to reduce the number of calculations

• Candidate pixels in left image– Hand picked pixels– Edges from left image– Every pixel

Stereopsis

• Correspondences + stereo geometry => depth => rendering from other locations

Results

Results

Ideas for Future Work

• Improve correlation quality metric

• Resolution pyramid for image: – start at low res when featureless surfaces are

smaller

• Resolution pyramid for pattern

• Inter-register several stereo pairs