Analysis of Motioncs332/ppt/sfm.pdf3 1-5 Using the rigidity constraint Ullman: Given 3 distinct...

6
1 CS332 Visual Processing Department of Computer Science Wellesley College Analysis of Motion Recovering 3-D structure from motion 1-2 Recovering 3D structure from motion

Transcript of Analysis of Motioncs332/ppt/sfm.pdf3 1-5 Using the rigidity constraint Ullman: Given 3 distinct...

  • 1

    CS332 Visual ProcessingDepartment of Computer ScienceWellesley College

    Analysis of Motion

    Recovering 3-D structurefrom motion

    1-2

    Recovering 3D structure from motion

  • 2

    1-3

    Ambiguity of 3D recovery

    birds’ eyeviews

    We need additional constraintto recover 3D structure uniquely

    “rigidity constraint”

    1-4

    Image projectionsZ

    X

    perspectiveprojection

    imageplane

    Z

    X

    orthographicprojection

    image plane

    (X, Y, Z) (X, Y)(X, Y, Z) (X/Z, Y/Z)

    • only relative depth

    • requires object rotation

    • only scaled depth

    • requires translation ofobserver relative to scene

  • 3

    1-5

    Using the rigidity constraint

    Ullman: Given 3 distinct orthographic viewsof 4 non-coplanar points in motion, if thereexists a rigid 3-D structure consistent withthese views, then this structure is unique.

    view 1 view 2 view 3

    2D positions of pointsin 3 views

    set ofequations to

    solve

    3D structureof points

    caveat: depth reversals

    1-6

    What is needed to compute a unique rigid 3D structure?

    5 points1 view

    5 points1-1/2 views

    7 points2 views

    4 points3 views

    orthographicprojection

    perspectiveprojection

    correspondence

    velocity field

    Sample results:

    *

    * Ullman

    Rigidity constraint alone is sufficient tocompute 3D structure from motion, BUT…

  • 4

    1-7

    Human recovery of 3D structure from motion:

    • Needs extended time to obtain an accurate3D structure, and derivation is not all-or-none

    • Can cope with significant deviations from rigidity

    • Integrates multiple sources of information

    These factors motivated the design ofUllman’s incremental rigidity scheme

    1-8

    Incremental Rigidity Scheme

    x

    y

    image

    (x1 y1 z1)

    (x2 y2 z2)

    (x3 y3 z3)

    (x1' y1' ??)

    (x2' y2' ??)

    (x3' y3' ??)

    depth: Z

    initially, Z=0at all points

    Find new 3D modelthat maximizes rigidity

    Compute new Zvalues that minimize

    change in 3Dstructure

  • 5

    1-9

    Bird’s eye view:de

    pth

    ximagecurrent model

    new imageFind new Zi that minimize

    Σ (Lij – lij)2/Lij3

    Lij

    lijZ

    1-10

    Incremental rigidity scheme: results

    bird’s eye views

    new 3D model afterevery 10° rotation

    true structure

    computed 3D model

    builds 3D modelincrementally over

    extended time

  • 6

    1-11

    Spontaneous depth reversal

    truestructure

    computed3D model

    1-12

    Human recovery of 3D structure from motion• 2-3 points sufficient

    • requires extended time to build up accuratepercept of 3-D structure

    Anderson & colleagues