deep learning for surface reconstruction · 2018. 3. 30. · the proposed deep learning som will be...

Post on 25-Mar-2021

2 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

SHAFAATUNNUR HASAN

SITI MARIYAM SHAMSUDDIN

UTM BIG DATA CENTRE,

UNIVERSITI TEKNOLOGI MALAYSIA,

81310 SKUDAI JOHOR

DEEP LEARNING FOR SURFACE RECONSTRUCTION

1.Universiti Teknologi Malaysia is an innovation-driven entrepreneurial Research University and a leading research-intensive university in engineering, science and technology ranked in the top 100 world ranking in engineering and technology.

2.It is located both in Kuala Lumpur, the capital city of Malaysia and Johor Bahru, the southern city in Iskandar Malaysia, which is a vibrant economic corridor in the south of Peninsular Malaysia

RESEARCH STRUCTURE IN UTM

Three (3) HiCoE

a. Wireless Communication Centre (WCC)

b. Advanced Membrane Technology Centre

c. Institute of Noise and Vibration (IKG)

Six (6) Research Institutes:

a. Ibnu Sina Institute for Scientific and Industrial Research (ISI-SIR)

b. Institute for Smart Infrastructures and Innovative Construction (ISIIC)

c. Institute for Vehicle Systems and Engineering (IVeSE)

d. Institute of Human CenteredEngineering (IHCE)

e. Institute of Future Energy (IFE)

f. Research Institute for Sustainable Environment (RISE)

Five (5) Research Alliances

a. Frontier Materials

b. Innovative Engineering

c. Health and Wellness

d. Resource Sustainability

e. Smart Digital Community

WHO WE ARE

UTM Big Data Centre - rooted from well-experienced researchers and practitionersfrom one of the research groups in UTM –Soft Computing Research Group (SCRG).With almost 20 years experiences in the fieldof Machine Learning, Pattern Recognition,Data Analytics, and Intelligent Graphicsmodelling, as well as in Big Data Analyticsand GPU-based Machine Learning.

PRIMARY FOCUSWhat is our expertise?

Data Science

Big Data

CO

PYR

IGH

T @

UTM

BIG

DA

TA C

ENTR

E

Big Data Science

CO

PYR

IGH

T @

UTM

Big

Dat

a C

entr

e

GPUMLiB-- AI as a Service

(AIaaS) @ UTM Big Data Centre

GPU MACHINE LEARNING LIBRARY

(GPUMLiB)- OPEN SOURCE LIBRARY

11

What we are presenting now

Deep Learning SOM for

Surface Reconstruction

.

The exponential growth of 3D objects, images, devices upto the Nth dimensional representation and constructionwill decrease the performance drastically. Thus, DeepLearning SOM algorithm is proposed in this study tooptimize the performance of 3D reconstruction andrepresentation.

Wai Pai Lee, Shafaatunnur Hasan, Siti Mariyam Shamsuddin and Noel Lopes. GPUMLib: Deep Learning SOM Library for Surface Reconstruction. International Journal of Advances in Soft Computing and its Application, 9, 2(2017), 1-16

What & Why Deep Learning for Surface Reconstruction

Where to implement Deep Learning SOM?

13

Surface Representation

Surface Reconstruction

Representing the point cloud data set into a viewable

state in a computer such as a computer vision object.

Including two categories: Explicit and Implicit.

Usually assumed as part of surface reconstruction.

The process of retrieving the point cloud data set

generated by a device.

Connecting the coordinates in point cloud data set which

are attached with the points information.

How to implement Deep-learning SOM in Close Surface Environment ?

How to implement Deep-learning SOM in Close Surface Environment ?

• Point Cloud Collection• PCL data set repositoryPhase 1

• Data Preprocessing• GPUMLib data representationPhase 2

• Deep Learning SOM• Deep Learning GPUMLib SOM Optimization

Phase 3

• Surface Representation• PCL ViewerPhase 4

SOM

SOM

How to implement Deep-learning SOM in Close Surface Environment ?GPUMLib Framework

SOM- GPUMLib RESEARCH DESIGN

GPUMLib Implementation

Host (CPU) and device (GPU) memory access framework Reduction framework

HostArray

DeviceArray

HostMatrix

DeviceMatri

x

CudaArray

MinIndex

SOM Implementation

Host (CPU) Device (GPU)

Read input data

Initialize weights

Compute distance and find

BMU

Update the weights

Display output

Termination

is satisfied?

Compute distanceComputeDistancesSOMkernel<<<…>>>

Find BMUMinSmallArrayIndex<<<…>>>

Copy value to hostUpdateHost()

Update the weightsUpdateWeightsSOMkernel<<<…>>>

Normalize the weightsNormalizeWeightsSOMkernel<<<…>>>

yes

no

How to implement Deep-learning SOM in Close Surface Environment ?

SOM-GPUMLib Research Design

Point Cloud with

Connectivity

Point Cloud without

Connectivity

Output

point based

on the final

mapping

from the

previous

layer

Output point

based on the

final weights

-Map size reducesas Iteration increases

How is the Architecture of Deep Learning SOM?The Proposed Deep-layer SOM

Original

Output Results

The Proposed Deep-learning SOM Surface Reconstruction Output

Single Layer SOM Deep-Layer SOM

Model Sphere Bunny1 Bunny2 Eagle Sphere Bunny1 Bunny2 Eagle

Points 422 8171 35947 796825 422 8171 35947 796825

Iteration 1000 1000 1000 500 1575 1575 1575 1575

MapX 10 25 25 100 20 40 40 100

MapY 20 40 50 200 40 60 60 200

Time (s) 4.84 124.528 619.12 52560 10.08 60.192 75.21 4392

CPU GPU

Model Sphere Bunny1 Bunny2 Eagle Sphere Bunny1 Bunny2 Eagle

Points 422 8171 35947 796825 422 8171 35947 796825

Iteration 1575 1575 1575 1575 1575 1575 1575 1575

MapX 20 40 40 100 20 40 40 100

MapY 40 60 60 200 40 60 60 200

Time (s) 4.30 64.70 87.32 9540 10.08 60.192 75.21 4392

Performance Comparison between CPU and GPU using Deep-Layer SOM

Deep-layer SOM Performance Comparison between Single Layer SOM and Deep-Layer SOM

Performance Comparison of Single Layer and Deep-layer SOM

21

The proposed Deep Learning SOMwill be an alternative solution fordeep optimization in searching, re-organizing and optimizing higherdimensional spaces for complexproblems, i.e., complex design andscenes. Our future work will bedeveloping mobile-based GPUMLib:Deep Learning SOM for optimizingcomplex objects, scenes and related.

AcknowledgementThe authors thank NVIDIA CORPORATION for the support in sponsoring the passes to GTC 2018; Malaysian Ministry of Higher Education (MOHE) for the financial support in conducting this project and Universiti Teknologi Malaysia for the R & D activities.

Conclusion & Future Direction

Short Demo

THANK YOU

&

Terima Kasih

THANK YOU

Contact : shafaatunnur@gmail.com; shafaatunnur@utm.mywebsite: bigdata.utm.myfacebook: @bigdatautm

top related