1 of 26

Low Rank Factorization-based for Deformable 3D Reconstruction

Discover how our proposed approach to 3D reconstruction tackles the challenges of non-rigid scenes, providing fast and robust results using factorization and neural radiance.

1

2 of 26

Task Force

ASRAR ALRUWAYQI

SHUBHAM TULSIANI

Adviser 

2

3 of 26

Reconstructing Non-Rigid Scenes in 3D 

Non-rigid reconstruction is a complex task of creating accurate depictions of objects that exhibit shape variability, such as dancing. Despite its challenges, scientists strive for the results to be as close to the original as possible. However, there are remaining problems that require further investigation.

3

4 of 26

Our drive is from:

NON-RIGID SCENES REFLECT THE TRUE NATURE OF OUR ENVIRONMENT.

FACTORIZATION ALLOWS US TO DECONSTRUCT COMPLEX DEFORMATIONS INTO MANAGEABLE COMPONENTS.

4

5 of 26

Problem Definition  

Our objective is to develop an effective deep learning model that can synthesize new views at an arbitrary time and explicitly encode a dynamic scene that is being captured by a monocular camera and moving non-rigidly from a sparse set of images. 

DYNAMIC SCENE FROM MONOCULAR CAMERA

GENRATE NEW VIEWS

5

6 of 26

Our Objectives

  • EFFICIENT ALGORITHM

Decompose tensor and use one mode to reconstruct 3d scene.

  • MANAGING DEFORMBEL OBJECT 

integrate between explicitly and implicitly representation. 

6

REF: NERF

7 of 26

Tensor Decompostion

Parallel Factors (PARAFAC), Canonical Decomposition (CANDECOMP), aka CP

7

THE CP DECOMPOSITION FACTORIZES A TENSOR INTO A SUM OF COMPONENT RANK-ONE TENSORS.

RER:BRETT W. BADER, TENSOR DECOMPOSITIONS AND APPLICATIONS∗

8 of 26

Tensor Decomposition  Existing works:

TensoRF ECCV 2022 (static)

Left: CP decomposition, which factorizes a tensor as a sum of vector outer products.

Right: vector-matrix decomposition.

8

RER: TESNORF

9 of 26

Tensor Decomposition  Existing works:

TensoRF ECCV 2022

9

RER: TESNORF

10 of 26

Tensor Decompostion

10

TENSORF 

OUR EXPERMINT :

ONE MODE

RER: TESNORF

11 of 26

TensoRF Training Process 

11

RER: TESNORF

12 of 26

Our Training Process  

12

13 of 26

13

14 of 26

Tensor Decomposition  Existing works:

Hex-Plane CVPR2023 (dynamic) 

14

15 of 26

Tensor Decomposition  Existing works:

Hex-Plane CVPR2023 (dynamic) 

15

16 of 26

Tensor Decomposition  Existing works:

K-Plane CVPR 2023 (dynamic) 

16

17 of 26

17

18 of 26

low-rank performans  

18

19 of 26

Failed Cases 

19

20 of 26

Dealing with Deformable Scene  Existing works

Nerfies ICCV 2021

20

21 of 26

Dealing with Deformable Scene  Existing works

Nerfies ICCV 2021

21

22 of 26

Dealing with Deformable Scene  Existing works

D-NerF CVPR 2020

22

23 of 26

Dealing with Deformable Scene  Existing works

D-NerF CVPR 2020

23

24 of 26

Next Experiment 

24

25 of 26

Timeline

LAUNCH

END

AUG 5

Brainstorming

SEPT 12

formulize solution 

SEP 17

integrate bawneen hexplane and Nerfies 

OCT 4

one mode experiments 

NOV 1

factorize deformable felid 

DEC 5

Optimization

25

26 of 26

Q/A