Scene Flow by Tracking in Intensity and Depth Data Julian Quiroga
Frederic Devernay
James Crowley
PRIMA team, INRIA Grenoble
[email protected]
June 12, 2012
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
1 / 20
Motivation
Surface Flow, Morpheo-INRIA 2011
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
Messing et al., ICCV 2009
June 12, 2012
2 / 20
Scene flow computation
Stereo or multiview: From several optical flows By using structure constrains and 2D/3D regularization
Basha et al., CVPR 2010
Simultaneously with 3D surface Tracking surfels (surface elements) Color and depth: Photometric constrains and 3D regularization Particle filtering in 3D Hadfield and Bowden, ICCV 2011
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
3 / 20
Our work Approach Sparse scene flow: we track small surface patches in the scene by using a pair of aligned intensity and depth images.
Model To constraint the scene flow in the image domain we assume a scene composed of rigidly moving 3D parts performing translation.
Framework By using the scene flow as parameter vector we extend the Lukas-Kanade approach to exploit both intensity and depth data.
Result We simultaneously solve for the scene flow and the image flow. Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
4 / 20
Presentation outline
Lucas-Kanade framework Motion model Locally rigid tracking approach Tracking in intensity and depth Experimentation Conclusion
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
5 / 20
Lucas-Kanade Framework The goal of Lucas-Kanade algorithm is to align a template image T (x) to an input image I(x). This problem can be stated as X P = arg min [I (W(x; P)) − T (x)]2 P
x
Assuming that an initial estimate of P is known, each optimization step finds ∆P which minimizes 2 X X ∂W 2 ∆P − T (x) [I (W(x; P + ∆P)) − T (x)] = I (W(x; P)) + ∇I ∂P x x Taking the partial derivative with respect to ∆P and solving it gives −1
∆P = H where H =
X ∂W T ∇I [I (W(x; P)) − T (x)] ∂P x
P ∂W T ∂W ∇I ∂P ∇I ∂P . x
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
6 / 20
Presentation outline
Lucas-Kanade framework Motion model Locally rigid tracking approach Tracking in intensity and depth Experimentation Conclusion
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
7 / 20
Motion model The instantaneous motion of a rigid surface point can be expressed as X0 = RX + V, The 3D motion of the surface generates the image flow given by X X − Y ΩZ + Z ΩY + VX 0 − fx u =x −x = Z − X ΩY + Y ΩX + V Z Z and 0
v =y −y =
Y − X ΩZ + Z ΩX + Vy Y − Z − X ΩY + Y ΩX + V Z Z
fy
Assuming that the inter-frame rotation is negligible , the image flow induced on a pixel x = (x, y ) by the 3D translation of the surface can be modeled as follows V 1 u 1 0 −x X VY = ∆(x; V) = 0 1 −y v Z VZ Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
8 / 20
Presentation outline
Lucas-Kanade framework Motion model Locally rigid tracking approach Tracking in intensity and depth Experimentation Conclusion
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
9 / 20
Locally rigid tracking approach Under brightness constancy assumption, points X at time t − 1 and X0 = X + V at time t are projected with the same intensity in the image ˆ (X + V) = I t−1 M(X) ˆ It M Considering a set of surface points S, the scene flow computation is stated as finding vector V = {VX , VY , VZ } which minimizes i2 Xh ˆ (X + V) − I t−1 M ˆ (X) It M X∈S
The imagen flow of each surface points is given by the warp function ˆ (X + V) = x + ∆(x; V) W(x; V) = M where ∆(x; V) is the proposed motion model. Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
10 / 20
Locally rigid tracking approach The problem can be formulated in the image domain as follows X V = arg min [I (W(x; V)) − T (x)]2 V
{x}∈s
Solution. Each element of the Jacobian is given by ∂W 1 1 0 −x = ∂V Z (x) 0 1 −y Over each iteration the Hessian matrix can be expressed as 2 Ix Ix Iy Ix IΣ X 1 Ix Iy Iy2 Iy IΣ H= 2 Z (x) {x} Ix IΣ Iy IΣ IΣ2 with IΣ = − xIx + yIy . Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
11 / 20
Presentation outline
Lucas-Kanade framework Motion model Locally rigid tracking approach Tracking in intensity and depth Experimentation Conclusion
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
12 / 20
Tracking in intensity and depth Under translation V the depth image must satisfy ˆ + V)) = Z t−1 (M(X)) ˆ Z t (M(X + VZ Therefore, we propose to formulate the scene flow computation by constraining V both in the intensity and depth: X
h i2 [I (W(x; V)) − T (x)]2 + λ Z (W(x; V)) − Tz (x) + D T V
{x}∈s
where λ = σI2 /σZ2 and D = (0, 0, 1) separates the Z component. Solution. X 1 Ix2 + λZx2 Ix Iy + λZx Zy Ix IΣ + λZx (ZΣ − 1) 2 2 Ix Iy + λZx Zy Iy + λZy Iy IΣ + λZy (ZΣ − 1) H= 2 2 Ix IΣ + λZx (ZΣ − 1) Iy IΣ + λZy (ZΣ − 1) IΣ + λ(ZΣ − 1)2 Z (x) {x} with IΣ = − xIx + yIy and ZΣ = − xZx + y Zy . Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
13 / 20
Presentation outline
Lucas-Kanade framework Motion model Locally rigid tracking approach Tracking in intensity and depth Experimentation Conclusion
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
14 / 20
Experimentation - Middlebury datastes
I1
I2
Z1
Details
Z2
Comparisons
Images : Teddy, Cones (2 and 6)
RT: proposed method (rigid traslation)
5 leves of PYR decomposition
RT0: proposed method with λ = 0
Window size: 11×11
KLT: KLT by Bouguet (OpenCV)
Image coverage: 85%
OFR : KLT with a robust norm Hug07 : Huguet and Devernay, ICCV 2007
Error measures Optical flow: RMSOF , AEEOF , RX
Bas10 : Basha et al., CVPR 2010
Scene flow: NRMSV , RX %
Had11 : Hadfield and Bowden, ICCV 2011
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
15 / 20
Experimentation - Middlebury datastes RMSOF R1.0 R5.0 AEEOF
RT 2.51 12.9 2.32 1.15
RT0 2.61 14.8 3.99 1.33
OFR 4.69 28.4 15.2 1.56
KLT 5.95 40.1 19.7 1.45
NRMSV R5% R20%
RMSOF R1.0 NRMSV R5%
RT Utex 3.11 39.8 10.9 28.5
RT0 55.9 28.8 9.68
OFR 68.4 37.9 17.5
KLT 82.0 38.1 19.1
Table 2: Errors in the scene flow.
Table 1: Errors in the optical flow. Tex 4.99 16.5 23.2 12.1
RT 11.1 17.1 4.97
DD 6.91 32.5 26.5 25.1
Tex 5.75 38.9 96.7 32.0
OFR Utex 7.20 58.8 202 51.5
DD 7.05 68.7 188 69.6
Table 3: Errors by regions. RMSOF (%) AEEOF
RT 5.70 2.47
Hug07 6.00 0.60
Bas10 2.96 0.70
Had11 0.10 5.03
Table 4: Scene flow comparison. Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
16 / 20
Experimentation - Kinect images
I1
Optical flow
I2
V Scene flow projection
Z1
Julian Quiroga (INRIA)
Vz
Scene Flow by Tracking in I+D
June 12, 2012
17 / 20
Presentation outline
Lucas-Kanade framework Motion model Locally rigid tracking approach Tracking in intensity and depth Experimentation Conclusion
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
18 / 20
Conclusion We have proposed a method to compute a sparse scene flow by using an aligned pair of intensity and depth images. Modeling the image flow as a function of the 3D motion field with help from the depth sensor allows the constraint of the scene flow of a small surface patch in the image domain. Combining intensity and depth data in a Lucas-Kanade framework we simultaneoulsy solve for the scene flow and image flow. This method is versatile and cand be used to generate more accurate trajectories or to define scene flow based descriptors.
Future work A criterion for selecting good regions to track Experimentation: action and gesture recognition
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
19 / 20
The End
Julian Quiroga (INRIA)
Scene Flow by Tracking in I+D
June 12, 2012
20 / 20