Pose Machines: Articulated Pose Estimation via Inference Machines
Model Based Pose Estimation
Transcript of Model Based Pose Estimation
![Page 1: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/1.jpg)
Model Based Pose Estimation
Gerard Pons-Moll and
Bodo Rosenhahn
Institute for Information Processing
![Page 2: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/2.jpg)
Title of the chapter
- Motion Analysis, Vicon (55+ Marker, 1000fps, Strobe-lights)
Overview
![Page 3: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/3.jpg)
Title of the chapter Overview • 3
Human pose estimation algorithms can be
classified in:
• Generative Models
- „explain the image“
• Discriminative Models
-„condition on the image“
![Page 4: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/4.jpg)
Title of the chapter
• Many degrees of freedom
• Highly Dynamic / Skinning/ Clothing / Outdoor
• Large variability and individuality of Motion patterns
http://www.google.de/images
Is it hard ?
![Page 5: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/5.jpg)
Title of the chapter
• Let us assume a model then:
• How to parameterize/represent the model ?
• How to optimize the model parameters ?
• What features are suited ?
http://www.google.de/images
Is it hard ?
![Page 6: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/6.jpg)
Title of the chapter Overview • 3
1) Kinematic parameterization - Rotation Matrices
- Euler Angles
- Quaternions
- Twists and Exponential maps
- Kinematic chains
2) Subject model - Geometric primitives
- Detailed Body Scans
- Human Shape models
3) Inference - Observation likelihood
- Local optimization
- Particle Based optimization
![Page 7: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/7.jpg)
Title of the chapter Human Motion • 3
![Page 8: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/8.jpg)
Title of the chapter
Motivated from robotics:
The human motion can be expressed via a „kinematic
chain“, a series of local rigid body motions (along the limbs).
Bregler et.al. CVPR-98
Kinematic Chains
The model parameters to optimize for are rigid
body motions.
How to model RBM ?
![Page 9: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/9.jpg)
Title of the chapter Parameterization 1) Pose configurations are represented with a
minimum number of parameters
2) Singularities can be avoided during
optimization
3) Easy computation of derivatives segment
positions and orientations w.r.t parameters
4) Human motion contrains such as
articulated motion are naturally described
5) Simple rules for concatenating motions
![Page 10: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/10.jpg)
Title of the chapter
A rigid body motion is an affine transformation that preserves distances
and orientations
Euclidean : (non linear)
Affine: (linear)
X =
0 B B @
a 1
a 2
a 3 1
1 C C A ! X ̀ =
µ R t
0 1
¶ X
R R T = I ; d e t ( R ) = 1
X =
0 @
a 1
a 2
a 3
1 A ! X ̀ = R X + t ; R 2 R 3 £ 3 ; t 2 R 3
Definition
![Page 11: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/11.jpg)
Title of the chapter Rotation Matrices • 3
S B
The columns of a rotation matrix are the
principal axis of one frame expressed relative
to another
![Page 12: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/12.jpg)
Title of the chapter 2 Views of Rotations • 3
Rotations can be interpreted either as
Coordinate
transformation
Relative motion
in time
![Page 13: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/13.jpg)
Title of the chapter Rotation matrix drawbacks
• Need for 9 numbers
• 6 additional constrains to ensure that the
matrix is orhtornormal
• Suboptimal for optimization
![Page 14: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/14.jpg)
Title of the chapter Euler Angles
• One of the most popular parameterizations
• Rotation is encoded as the successive
rotations about three principal axis
• Only 3 parameters to encode a rotation
• Derivatives easy to compute
![Page 15: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/15.jpg)
Title of the chapter Euler Angles
S
![Page 16: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/16.jpg)
Title of the chapter Euler Angles: confusion
Careful: Euler angles are a typical source of
confusion
When using Euler angles 2 things have to be
specified
1) Convention: X-Y-Z, Z-Y-X, Z-Y-Z …
2) Rotations about the static spatial frame or
the moving body frame
![Page 17: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/17.jpg)
Title of the chapter Euler Angles: drawbacks I
• Gimbal lock: When two of the axis align one
degree of freedom is lost !
• Parameterization is not unique
• Lots of conventions for Euler angles
![Page 18: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/18.jpg)
Title of the chapter Quaternions
• A quaternion has 4 components:
• They generalize complex numbers
with additional properties
• Unit length quaternions can be used to carry
out rotations. The set they form is called
![Page 19: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/19.jpg)
Title of the chapter Quaternions
• Quaternions can also be interpreted as a
scalar plus a 3-vector
Where
![Page 20: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/20.jpg)
Title of the chapter Quaternions
• Rotations can be carried away directly in
parameter space via the quaternion product:
- Concatenation of rotations:
- If we want to rotate a vector
where is the quat conjugate
![Page 21: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/21.jpg)
Title of the chapter Quaternions
Quaternions have no singularities
Derivatives exist and are linearly
independent
Quaternion product allows to perform
rotations
But all this comes at the expense of using
4 numbers instead of 3
- Enforce quadratic constrain
![Page 22: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/22.jpg)
Title of the chapter Axis-angle
For human motion modeling it is often needed
to specify the axis of rotation of a joint
Any rotation about the origin can be
expressed in terms of the axis of rotation
and the angle of rotation with the
exponential map
![Page 23: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/23.jpg)
Title of the chapter Lie Groups / Lie Algebras
Lie Group
M Lie algebra
exp
0M
Definition: A group is an n-dimensional Lie-group,
if the set of its elements can be represented as a
continuously differentiable manifold of dimension n,
on which the group product and inverse are
continuously differentiable functions as well
![Page 24: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/24.jpg)
Title of the chapter Lie Groups / Lie Algebras
= µ
µ ¡ s i n ( Á ) ¡ c o s ( Á ) c o s ( Á ) ¡ s i n ( Á )
¶ j 0
= µ
µ 0 ¡ 1 1 0
¶ = µ ^ !
s o ( 2 ) = f A 2 R 2 £ 2 j A = ¡ A T g
s o ( 2 ) =
µ c o s ( Á ) ¡ s i n ( Á ) s i n ( Á ) c o s ( Á )
¶ ; µ @
µ c o s ( Á ) ¡ s i n ( Á ) s i n ( Á ) c o s ( Á )
¶ j 0
Example:
(1)
(1) Is a time invariant linear differential equation which may be integrated to give:
If a body rotates at constant velocity about an axis, the velocity can be written as
q ( t ) = e x p ( ̂ ! t ) q ( 0 )
µ 0 ¡ 1 1 0
¶ µ 1
0
¶ =
µ 0
1
¶ ;
µ 0 ¡ 1 1 0
¶ µ 0
1
¶ =
µ ¡ 1 0
¶ _ q ( t ) = ^ ! q ( t )
![Page 25: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/25.jpg)
Title of the chapter Axis-angle
Given a vector the skew symetric matrix is
It is the matrix form of the cross-product:
You will also find
it as
![Page 26: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/26.jpg)
Title of the chapter Exponential map
Proof: exponential map
If we rotate units of time
![Page 27: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/27.jpg)
Title of the chapter Exponential map
Exploiting the properties of skew symetric
matrices
Rodriguez formula
Closed form!
![Page 28: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/28.jpg)
Title of the chapter Twists
What about translation ?
The twist coordinates are defined as
And the twist is defined as
![Page 29: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/29.jpg)
Title of the chapter Exponential map
The rigid body motion can be computed in
closed form as well
From the following formula
![Page 30: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/30.jpg)
Title of the chapter Ranking
Number of
parameters
Singularities Human
constraints
Concatenate
motion
Optimization
(derivatives)
Twists Quaternions Twists Quaternions Twists
Euler Angles Twists Quaternions Twists Euler Angles
Quaternions Euler Angles Euler Angles Euler Angles Quaternions
![Page 31: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/31.jpg)
Title of the chapter Overview • 3
1) Kinematic parameterization - Rotation Matrices
- Euler Angles
- Quaternions
- Twists and Exponential maps
- Kinematic chains
2) Subject model - Geometric primitives
- Detailed Body Scans
- Human Shape models
3) Inference - Observation likelihood
- Local optimization
- Particle Based optimization
![Page 32: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/32.jpg)
Title of the chapter Articulation
S
B
In a rest position we have
![Page 33: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/33.jpg)
Title of the chapter Articulation
S
![Page 34: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/34.jpg)
Title of the chapter Articulation
S
![Page 35: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/35.jpg)
Title of the chapter Articulation
S
The coordinates of the point in the
spatial frame
![Page 36: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/36.jpg)
Title of the chapter Product of exponentials
Product of exponentials formula
is the mapping from coordinate B
to coordiante S
BUT IS NOT the mapping from
segment i+1 to segment i.
Think of simply as the relative
motion of that joint
![Page 37: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/37.jpg)
Title of the chapter Inverse Kinematics
£Thatisatest
Supose we want to find the angles to reach
a specific goal
![Page 38: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/38.jpg)
Title of the chapter Inverse Kinematics
£Thatisatest
Supose we want to find the angles to reach
a specific goal
• The problem is non-linear
• Linearize with the
articulated Jacobian
![Page 39: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/39.jpg)
Title of the chapter Articulated Jacobian
The Jacobian using twists is extremely simple
and easy to compute
1) Every column corresponds to the
contribution of i-th joint to the end-effector
motion
2) Maps an increment of joint angles to the
end-effector twist
![Page 40: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/40.jpg)
Title of the chapter Articulated Jacobian
B
Intuition: Linear combination of twists
![Page 41: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/41.jpg)
Title of the chapter Articulated Jacobian
Intuition: Linear combination of twists
![Page 42: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/42.jpg)
Title of the chapter Articulated Jacobian
Intuition: Linear combination of twists
![Page 43: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/43.jpg)
Title of the chapter Pose Parameters
Pose parameters: root + joint
angles
![Page 44: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/44.jpg)
Title of the chapter Pose Jacobian
Maps increments in the pose parameters to
increments in end-effector position
6 columns of
Root
N columns for
one per joint
![Page 45: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/45.jpg)
Title of the chapter Overview • 3
1) Kinematic parameterization - Rotation Matrices
- Euler Angles
- Quaternions
- Twists and Exponential maps
- Kinematic chains
2) Subject model - Geometric primitives
- Detailed Body Scans
- Human Shape models
3) Inference - Observation likelihood
- Local optimization
- Particle Based optimization
![Page 46: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/46.jpg)
Title of the chapter Geometric primitives 2D 3D
Kjellström et.al.
Sigal et.al.
Kehl and Van Gool
Sminchisescu and Triggs
Felzenszwalb et.al
Ramanan et.al.
Andriluka et.al.
Plaenkers and Fua
Cylinders Ellipsoids Gaussian
Blobs
![Page 47: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/47.jpg)
Title of the chapter Detailed models
Pons-Moll et.al.
Rosehnahn et.al.
Hasler et.al.
Aguiar et.al.
Gall et.al
Cagniart et.al.
Rigged Subject Scan Free form Surface
~ 30 DoF
- Kinematic model
- > 1000 DoF
- with ++ constrains
![Page 48: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/48.jpg)
Title of the chapter Model Rigging
Skinning Animate
Skeletton
Non-rigid
registration
![Page 49: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/49.jpg)
Title of the chapter Fit a template
Correspondences of pairwise points
with similar local regions and similar
geodesic distances
Template Point cloud
Anguelov et.al
Loopy belief
propagation
![Page 50: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/50.jpg)
Title of the chapter Non-rigid registration
Distance term Smoothness term
Least squares
3x4 affine matrix
![Page 51: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/51.jpg)
Title of the chapter Shape and Pose models
Ballan et.al
Guan et.al
Hasler et.al SCAPE
Hasler et.al
Anguelov et.al
Infer model parameters
from images
Learn a PCA model of shape
![Page 52: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/52.jpg)
Title of the chapter Overview • 3
1) Kinematic parameterization - Rotation Matrices
- Euler Angles
- Quaternions
- Twists and Exponential maps
- Kinematic chains
2) Subject model - Geometric primitives
- Detailed Body Scans
- Human Shape models
3) Inference - Observation likelihood
- Local optimization
- Particle Based optimization
![Page 53: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/53.jpg)
Title of the chapter Inference
Optimization Bayesian models
Generative models
Posterior Likelihood Prior
Map of Approx. with
weighted samples
![Page 54: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/54.jpg)
Title of the chapter Optimization
Extract features Predict and
match
Optimize
![Page 55: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/55.jpg)
Title of the chapter Image features
• Silhouettes
• Edges
• Distance transforms
• SIFT
• Optic flow
• Appearance
• …
Any feature that can be predicted from the
model and is fast to compute
![Page 56: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/56.jpg)
Title of the chapter Optimization
O
Project model
Image features
Optimize Model-Image
![Page 57: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/57.jpg)
Title of the chapter Matching
1) Look along normal model
contour directions 2) Discretize and match
TIPS:
1) Match image to model and model to image
2) Careful removing outliers
![Page 58: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/58.jpg)
Title of the chapter Least squares
Model
predictions
Image
observations
Assuming Gaussian distribution it is
equivalent to a MAP estimate
![Page 59: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/59.jpg)
Title of the chapter Least squares
Express the problem in vector form
Residual for match 1
![Page 60: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/60.jpg)
Title of the chapter Local Optimization
Take a step in that
direction
Gradient ~Hessian
![Page 61: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/61.jpg)
Title of the chapter Jacobian
O
C
1) Pose
2) World-camera
transformation
3) Projection
1 2 3
![Page 62: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/62.jpg)
Title of the chapter Other likelihoods
2D-3D error
point-to-line distance
3D-3D error
point-to-point distance
![Page 63: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/63.jpg)
Title of the chapter Distance transforms
inconsistent consistent
1) Push model inside silhouette
2) Force the model to explain the image
Distance transform + overlapp term
Sminchisescu F & G 2001
![Page 64: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/64.jpg)
Title of the chapter Region based
Region-based Optical flow
Bregler and Malik
Parameterize flow with human
motion model Use model as region mask Q that
separates foreground from
background
Rosenhahn et.al.
![Page 65: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/65.jpg)
Title of the chapter Local optimization
It is fast and accurate
Prone to local minima
Requires initialization
Matching cost is ambiguous
Single hypothesis propagated
![Page 66: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/66.jpg)
Title of the chapter Overview • 3
1) Kinematic parameterization - Rotation Matrices
- Euler Angles
- Quaternions
- Twists and Exponential maps
- Kinematic chains
2) Subject model - Geometric primitives
- Detailed Body Scans
- Human Shape models
3) Inference - Observation likelihood
- Local optimization
- Particle Based inference
![Page 67: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/67.jpg)
Title of the chapter Particle Filter
• Once I know , is independent on previous measurements
Image observations at t
State space, pose parameters at t
First order Markov
process
• Once I know the state, the new measurement becomes
independent on the others
![Page 68: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/68.jpg)
Title of the chapter Particle Filter
Distribution approximated with a set of
weighted samples
![Page 69: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/69.jpg)
Title of the chapter Particle Filter
Posterior t-1
Posterior t
Temporal Dynamics
Diffusion
![Page 70: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/70.jpg)
Title of the chapter Particle Filter
Condensation, Isaard and Blake 1996
![Page 71: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/71.jpg)
Title of the chapter Resampling
Resample with
probability equal to
the weights
![Page 72: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/72.jpg)
Title of the chapter Sampling
sample weight
weight
![Page 73: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/73.jpg)
Title of the chapter Problems
Observation likelihood is highly multimodal !
1) Multiple optima
2) Huge search space
Video from Sminchisescu and Triggs
![Page 74: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/74.jpg)
Title of the chapter Annealed Particle Filter
Iteratively evaluate smooth versions of
Particles reduced by
a factor >10
Less prone to local
optima
Not as robust as
Bayesian
Still computationaly
expensive
Deutscher et.al.
Gall et.al.
weight
diffuse
resample
![Page 75: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/75.jpg)
Title of the chapter Efficient sampling I
Likelihood levelsets
Samples
Localy optimize every
sample of MCMC
Cho and Fleet
Hybrid MCMC
![Page 76: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/76.jpg)
Title of the chapter Efficient sampling II
Covariance Scaled Sampling
Scatter particles along cost
function valley
Explore high dimensional
space more efficiently
Dedicates some particles to
explore globaly
Sminchisescu and Triggs
![Page 77: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/77.jpg)
Title of the chapter Discussion
Generative modeling:
- Need to model Kinematics
- Need to model Shape
- Need to model Observation Likelihood
- Texture
- Ilumination
- ufff lots of work so …
IS THIS THE END OF GENERATIVE ?
![Page 78: Model Based Pose Estimation](https://reader031.fdocuments.in/reader031/viewer/2022012409/616a454a11a7b741a350ad2f/html5/thumbnails/78.jpg)
Title of the chapter End of Generative ?
Well, depends on the application…
In totaly uncontrolled scenarios will never work!
But the accuracy is still higher and they
generalize to complex motions better than
discriminative approaches
Useful as refinement stage coupled with
discriminative initialization