46
Dense Variational Reconstruction of Non-Rigid Surfaces from Monocular Video Ravi Garg Anastasios Roussos * Lourdes Agapito * Queen Mary, University of London * Now at UCL Before This Paper 1 / 17

Ln l.agapito

Embed Size (px)

Citation preview

Page 1: Ln l.agapito

Dense Variational Reconstruction of Non-RigidSurfaces from Monocular Video

Ravi Garg Anastasios Roussos∗ Lourdes Agapito∗

Queen Mary, University of London∗Now at UCL

Before This Paper

1 / 17

Page 2: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

Goal: dense 3D reconstruction for every frame.2 / 17

Page 3: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

Goal: dense 3D reconstruction for every frame.2 / 17

Page 4: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

Goal: dense 3D reconstruction for every frame.2 / 17

Page 5: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

Goal: dense 3D reconstruction for every frame.2 / 17

Page 6: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

Goal: dense 3D reconstruction for every frame.2 / 17

Page 7: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

Goal: dense 3D reconstruction for every frame.

Goal: dense 3D reconstruction for every frame.

NO additional sensors.

NO pre-trained shape models.

NO surface template.

2 / 17

Page 8: Ln l.agapito

Dense Non Rigid Structure from Motion

Input: monocular sequence of non-rigid surface.

...

NRSfM: ill posed problem

Goal: dense 3D reconstruction for every frame.

Goal: dense 3D reconstruction for every frame.

NO additional sensors.

NO pre-trained shape models.

NO surface template.

2 / 17

Page 9: Ln l.agapito

Traditional Sparse Non Rigid Structure from Motion

...

3 / 17

Page 10: Ln l.agapito

Traditional Sparse Non Rigid Structure from Motion

...Feature Tracking

...

3 / 17

Page 11: Ln l.agapito

Traditional Sparse Non Rigid Structure from Motion

...Feature Tracking

...

...3D ShapeInference

3 / 17

Page 12: Ln l.agapito

Traditional Sparse Non Rigid Structure from Motion

Priors

...Feature Tracking

... +

...3D ShapeInference

3 / 17

Page 13: Ln l.agapito

Low Rank Prior for NRSfM

Shape Space

(Bregler, Hertzmann, Biermann, Recovering non-rigid 3D shape from image streams CVPR’00.)

Bregler et al. CVPR’00, Brand CVPR’01, Xiao et al. IJCV’06, Torresani et al. PAMI’08, Akhter et al.CVPR’09, Bartoli et al. CVPR2008, Paladini et al. IJCV’12,Dai et al. CVPR’12...

4 / 17

Page 14: Ln l.agapito

Low Rank Prior for NRSfM

Shape Space

(Bregler, Hertzmann, Biermann, Recovering non-rigid 3D shape from image streams CVPR’00.)

(Park et al. ECCV’10)

4 / 17

Page 15: Ln l.agapito

Inspiration from Dense Rigid Reconstruction

(Newcombe, Lovegrove, Davison, DTAM: Dense Tracking and Mapping in Real-Time, ICCV’11)

5 / 17

Page 16: Ln l.agapito

Inspiration from Dense Rigid Reconstruction

(Newcombe, Lovegrove, Davison, DTAM: Dense Tracking and Mapping in Real-Time, ICCV’11)

Key features

Variational approach.

Use of smoothness priors.

Per pixel reconstruction.

Scalable and GPU friendly algorithm.

5 / 17

Page 17: Ln l.agapito

Leap from sparse to dense NRSfM

Sparse

Dai et al. CVPR’12

Dense

This work

6 / 17

Page 18: Ln l.agapito

Leap from sparse to dense NRSfM

Sparse

Dai et al. CVPR’12

Dense

This work

We take the best of both worlds:

Low rank prior from sparse non rigid SfM.Variational framework from dense rigid SfM.

6 / 17

Page 19: Ln l.agapito

Leap from sparse to dense NRSfM

Sparse

Dai et al. CVPR’12

Dense

This work

We take the best of both worlds:

Low rank prior from sparse non rigid SfM.Variational framework from dense rigid SfM.

Our contribution

First variational formulation to dense NRSfM.

Scalable algorithm which can be ported on GPU.

6 / 17

Page 20: Ln l.agapito

Our Approach in a Nutshell

...

7 / 17

Page 21: Ln l.agapito

Our Approach in a Nutshell

...

...

Step 1: DenseVideo Registration ...

7 / 17

Page 22: Ln l.agapito

Our Approach in a Nutshell

Step 2: DenseShape Inference ...

...

...

Step 1: DenseVideo Registration ...

7 / 17

Page 23: Ln l.agapito

Our Approach in a Nutshell

Step 2: DenseShape Inference ...

...

...

Step 1: DenseVideo Registration ...

Priors +

7 / 17

Page 24: Ln l.agapito

Our Approach in a Nutshell

Step 2: DenseShape Inference ...

...

...

Step 1: DenseVideo Registration ...

Low rank.Spatial smoothness.+

7 / 17

Page 25: Ln l.agapito

Our Approach in a Nutshell

...

... Low rank.Spatial smoothness.

Step 1: DenseVideo Registration ...

+

Garg, Roussos, Agapito, A variational approach to video registration with subspace constraints, IJCV’13.7 / 17

Page 26: Ln l.agapito

Our Approach in a Nutshell

Step 2: DenseShape Inference ...

... Low rank.Spatial smoothness.+

7 / 17

Page 27: Ln l.agapito

Orthographic Projection Model

8 / 17

Page 28: Ln l.agapito

Orthographic Projection Model

8 / 17

Page 29: Ln l.agapito

Orthographic Projection Model

8 / 17

Page 30: Ln l.agapito

Orthographic Projection Model

8 / 17

Page 31: Ln l.agapito

Orthographic Projection Model

W = RS

8 / 17

Page 32: Ln l.agapito

Energy Minimisation Approach to NRSfM

Formulation of a single unified energy to estimate:

Orthographic projection matrices

3D shapes for all the frames

E(

R , S)

= λ Edata

(R,S

)+ Ereg

(S)

+ τ Etrace

(S)

reprojection error over all frames

spatial smoothness prior on 3D shapes

low rank prior on 3D shapes

9 / 17

Page 33: Ln l.agapito

Energy Minimisation Approach to NRSfM

Formulation of a single unified energy to estimate:

Orthographic projection matrices

3D shapes for all the frames

E(

R , S)

= λ Edata

(R,S

)+ Ereg

(S)

+ τ Etrace

(S)

reprojection error over all frames

spatial smoothness prior on 3D shapes

low rank prior on 3D shapes

9 / 17

Page 34: Ln l.agapito

Energy Minimisation Approach to NRSfM

Formulation of a single unified energy to estimate:

Orthographic projection matrices

3D shapes for all the frames

E(

R , S)

= λ Edata

(R,S

)+ Ereg

(S)

+ τ Etrace

(S)

reprojection error over all frames

spatial smoothness prior on 3D shapes

low rank prior on 3D shapes

9 / 17

Page 35: Ln l.agapito

Energy Minimisation Approach to NRSfM

Formulation of a single unified energy to estimate:

Orthographic projection matrices

3D shapes for all the frames

E(

R , S)

= λ Edata

(R,S

)+ Ereg

(S)

+ τ Etrace

(S)

reprojection error over all frames

spatial smoothness prior on 3D shapes

low rank prior on 3D shapes

9 / 17

Page 36: Ln l.agapito

Reprojection ErrorE

`R,S

´= λEdata

`R,S

´+ Ereg

`S

´+ τEtrace

`S

´Edata (R,S) = ‖W − RS‖2F

10 / 17

Page 37: Ln l.agapito

Spatial Smoothness PriorE

`R,S

´= λEdata

`R,S

´+ Ereg

`S

´+ τEtrace

`S

´Ereg

(S)

=∑

i

TV (Si)

−−−−−−→

Without regularisation With regularisation11 / 17

Page 38: Ln l.agapito

Low Rank PriorE

`R,S

´= λEdata

`R,S

´+ Ereg

`S

´+ τEtrace

`S

´

Etrace

(S)

= ‖S‖∗ =∑

i

σi(S)

lies in−−−−→ span

K � F

Angst et al. ECCV’12, Dai et al. CVPR’12, Angst et al. ICCV’11, Dai et al. ECCV’10

12 / 17

Page 39: Ln l.agapito

Minimisation of E(R,S

)

minR,Sλ ‖W − RS‖2F︸ ︷︷ ︸

Reprojectionerror

+∑i

TV (Si)︸ ︷︷ ︸Smoothness

prior

+ τ ‖S‖∗︸︷︷︸Low rank

prior

13 / 17

Page 40: Ln l.agapito

Minimisation of E(R,S

)

minR,Sλ ‖W − RS‖2F︸ ︷︷ ︸

Reprojectionerror

+∑i

TV (Si)︸ ︷︷ ︸Smoothness

prior

+ τ ‖S‖∗︸︷︷︸Low rank

prior

Our Algorithm

Initialize R and S using rigid factorisation.

Minimize energy via alternation:

Step 1: Rotation estimation.Step 2: Shape estimation.

Efficient and highly parallelizable algorithm → GPU-friendly

13 / 17

Page 41: Ln l.agapito

Minimisation of E(R,S

)

minRλ ‖W − RS‖2F︸ ︷︷ ︸

Reprojectionerror

+∑i

TV (Si)︸ ︷︷ ︸Smoothness

prior

+ τ ‖S‖∗︸︷︷︸Low rank

prior

Step 1: Rotation estimation

Robust estimation by using dense data.

Solved via Levenberg-Marquardt algorithm.

Rotations are parametrised as quaternions.

13 / 17

Page 42: Ln l.agapito

Minimisation of E(R,S

)

minSλ ‖W − RS‖2F︸ ︷︷ ︸

Reprojectionerror

+∑i

TV (Si)︸ ︷︷ ︸Smoothness

prior

+ τ ‖S‖∗︸︷︷︸Low rank

prior

Step 2: Shape estimation

Convex sub-problem.

Optimisation via alternation between:

Per frame shape refinement: using primal dual algorithmEnforcing low rank: using soft impute algorithm.

13 / 17

Page 43: Ln l.agapito

Results on real sequences

14 / 17

Page 44: Ln l.agapito

Quantitative Evaluation

Average RMS 3D reconstruction errors.

Sequence TB MP Ours Ours(τ = 0)

Non-smooth rotations 4.50% 5.13% 2.60% 3.32%Smooth rotations 6.61% 5.81% 2.81% 3.89%

- TB: Akhter et al, Trajectory space: A dual representation for non-rigid structure from motion, PAMI’11.

- MP: Paladini et al, Optimal metric projections for deformable and articulated structure-from-motion,IJCV’12.

15 / 17

Page 45: Ln l.agapito

Conclusions and future work

Conclusions:

First dense, template-free approach to Non-rigid Structurefrom Motion.Unified energy minimization for both rotation and shapeestimation.Combination of low-rank and spatial regularization prior.Using variational methods, we can do much more withmonocular sequences that one could expect

Future work:

Direct estimation from pixel intensitiesTowards real-time: online formulationOcclusion modelling

16 / 17

Page 46: Ln l.agapito

Thank You for Your Attention!

For more details and data

Visit: www.eecs.qmul.ac.uk/~rgargOR Come to our poster!

Questions?

Authors thank

Chris Russell and Sara Vicente for valuable discussions.17 / 17