1 technische universität münchen, 2 siemens corporate ......miroslava slavcheva1,2, maximilian...

1
KillingFusion: Non-rigid 3D Reconstruction without Correspondences Miroslava Slavcheva 1,2 , Maximilian Baust 1 , Daniel Cremers 1 , Slobodan Ilic 1,2 1 Technische Universität München, 2 Siemens Corporate Technology Task: real-time non-rigid 3D reconstruction Multiple objects Input: single, noisy depth stream. #0 #50 #85 #94 Warped live frames. Output: canonical- pose model. Quantitative evaluation dataset Reconstruction error [mm] VolumeDeform [Innmann16] KillingFusion [ours] Duck 5.362 3.896 Snoopy 4.205 3.543 Killing Fusion Volume Deform Live frame Topology changes & fast motion Results Objectives Topological changes. Rapid inter-frame motion. Handle only contrived motion. [all] Processing takes hours. [Dou15] Intermittent conversion from SDF to mesh for cor- respondence estimation. [Newcombe15, Innmann16] Need to estimate a 6D deformation field. [all] Limitations of related approaches References [Dou15] 3D Scanning Deformable Objects with a Single RGBD Sensor, CVPR 2015. [Newcombe15] DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-Time, CVPR 2015. [Innmann16] VolumeDeform: Real-time Volumetric Non-rigid Reconstruction, ECCV 2016. Contributions Full non-rigid framework based on a single data representation the signed distance field (SDF). Inherently handles topological changes. No correspondence estimation. No intermittent mesh conversion. SDF mesh Proposed approach Incremental variational estimation of a 3D defor- mation field Ψ = (u,v,w) from the projective SDF Φ n of the current frame towards canonical-pose model Φ global . Data term: alignment. Level set regularizer: geometric correctness. Motion regularizer: rigidity. Killing damped K set level S data E E E E 2 , , , , , , 2 1 z y x global n data z y x w z v y u x E 2 , , 1 , , 2 1 z y x n set level w z v y u x E J vec J vec J vec J vec E T T z y x T Killing damped , , Scene flow: Killing vector field (divergence-free): J vec J vec J vec J vec J J T T z y x T z y x F T , , , , 2 2 1 z y x z y x T J vec J vec w v u , , , , 2 2 2 Other datasets Effects of energy components All terms w S = 0: no level set regularizer w K = 0: no motion regularizer γ = 0: no Killing component γ = 1: pure Killing condition Full 3D Limitations No correspondences cannot propagate texture. Regular voxel grid real- time limited to ~80 3 voxels.

Upload: others

Post on 23-Mar-2021

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 1 Technische Universität München, 2 Siemens Corporate ......Miroslava Slavcheva1,2, Maximilian Baust1, Daniel Cremers1, Slobodan Ilic1,2 1 Technische Universität München, 2 Siemens

KillingFusion: Non-rigid 3D Reconstruction without Correspondences

Miroslava Slavcheva1,2, Maximilian Baust1, Daniel Cremers1, Slobodan Ilic1,2

1 Technische Universität München, 2 Siemens Corporate Technology

Task: real-time non-rigid 3D reconstruction

Multiple objectsInput: single, noisy

depth stream.

#0 #50 #85 #94

Warped

live frames.

Output: canonical-

pose model.

Quantitative evaluation dataset

Reconstruction

error [mm]

VolumeDeform

[Innmann16]

KillingFusion

[ours]

Duck 5.362 3.896

Snoopy 4.205 3.543

Kill

ing

Fusio

nV

olu

me

Defo

rm

Liv

e

fra

me

Topology changes & fast motion

Results

Objectives

• Topological changes. • Rapid inter-frame motion.

• Handle only contrived motion. [all]

• Processing takes hours. [Dou15]

• Intermittent conversion from SDF to mesh for cor-

respondence estimation. [Newcombe15, Innmann16]

• Need to estimate a 6D deformation field. [all]

Limitations of related approaches

References[Dou15] 3D Scanning Deformable Objects

with a Single RGBD Sensor, CVPR 2015.

[Newcombe15] DynamicFusion:

Reconstruction and Tracking of Non-rigid

Scenes in Real-Time, CVPR 2015.

[Innmann16] VolumeDeform: Real-time

Volumetric Non-rigid Reconstruction, ECCV

2016.

Contributions

• Full non-rigid framework based on a single data

representation – the signed distance field (SDF).

• Inherently handles topological changes.

• No correspondence estimation.

• No intermittent mesh conversion.

SDF mesh

Proposed approach

Incremental variational estimation of a 3D defor-

mation field Ψ = (u,v,w) from the projective SDF Φn of

the current frame towards canonical-pose model Φglobal.

• Data term: alignment.

• Level set regularizer: geometric correctness.

• Motion regularizer: rigidity.

KillingdampedK

setlevelSdata EEEE

2,,

,,,,2

1

zyx

globalndata zyxwzvyuxE

2,,

1,,2

1

zyx

nsetlevel wzvyuxE

JvecJvecJvecJvecETT

zyx

T

Killingdamped

,,

Scene flow:

Killing vector field

(divergence-free): JvecJvecJvecJvecJJ

TT

zyx

T

zyxF

T ,,,,

2

2

1

zyx zyx

TJvecJvecwvu

,, ,,

222

Other datasets

Effects of energy components

All terms wS = 0:

no level set

regularizer

wK = 0:

no motion

regularizer

γ = 0:

no Killing

component

γ = 1:

pure Killing

condition

Full 3D

Limitations

• No correspondences –

cannot propagate texture.

• Regular voxel grid – real-

time limited to ~803 voxels.