A Motion-Aware Approach to Continuous Retrieval of 3D Objects

(ICDE 2008)Mohammed Eunus Ali

Rui Zhang

Egemen Tanin

Lars Kulik

Department of Computer Science and Software EngineeringUniversity of Melbourne, Australia

Outline

Applications and problem

Our Motion-Aware approach

Data representation and retrieval

Buffer

Index

Experiments

Conclusion

A smart phone to see the interior of restaurants

Applications

Emerging more complex applications, e.g., tours using augmented reality

A rescue officer can see the structure of a building even if the building is on fire and filled with smoke

Problem

Continuous retrieval of 3D objects in a window

Model: client-server

Bottle neck: bandwidth, especially when the view is moving fast

Observation

QtQt+1

Qt+2

Qt+3

Qt+4

Qt+5

Qt+6

Speed

Speed

Speed

QtQt+1

Qt+2

Qt+3

Qt+4

Qt+5

Qt+6

A continuous query from a mobile client

The details can be determined using the client’s motion

Motion-Aware Approach

Motion-aware data retrieval (overall)

representing 3D objects in multiple resolutions (wavelets)

only retrieving necessary resolution (speed)

incremental retrieval (windows, resolutions)

Motion-aware buffer management (client)

prefetching

caching

Index for 3D objects in wavelets (server)

Base meshBase meshBase meshBase mesh

Progressively including detailsProgressively including detailsProgressively including detailsProgressively including details

Multi-resolution Representations

Figure: http://research.microsoft.com/~hoppe/

Example Wavelet Decomposition

Base Mesh (M0) Mesh (M1)M0 M1

v1v2

v3

v1v2

v3

v1v2

v3

v4

v5v6

v4

v5v6

Wavelet coefficient, d4 = v4 – (v1+v2)/2 = v4 – v′4

v′4

v′5v′6

Example Wavelet Decomposition

Mesh (M1)

v1v2

v3

v4

v5v6

Mesh (M2)

v1v2

v3

v4

v5v6

v’12

v’14

v’15

v’10

v’13

v’11

v’8

v’7

v’9

v1v2

v3

v4

v5v6

v12

v14

v15

v10

v13

v11

v8

v7

v9

M1 M2

1 2

3

4

56

Incremental Retrieval (window)

Q t-1

Q t

A B

CD

A’ B’

C’D’

E

F

G

Data Retrieval in Multiple Resolutions

Ot Qt Qt-1

Nt Qt - Qt-1

rt MapSpeedToResolution (st )

If ( Ot ) then

If (rt > rt-1) thenR Retrieve( {(Ot, rt-1, rt ), (Nt , 0, rt )} )

R Retrieve( {(Nt, 0, rt )} )

R Retrieve( {(Qt , 0, rt )} )

else

else

Algorithm: ContinuousRetrieval

QtQtQt-1Qt-1

Qt-1

Qt-1

Qt-1

Qt

Qt

Qt

Qt

Qt

Qt-1

rt

rt-1

Motion-Aware Buffer Management

• We have a high-latency environment with decent computational capacity

• Cache and pre-fetch objects that are very likely to be retrieved along the path of a client

• Kalman-Filter is used in target tracking

Prediction

Qt

Qt+10.5

0.2

0.3

Buffer: Given probabilities to move in one dimension to two directions

Find: nopt

Will maximize the Average Residency Time!

Buffer Assignment in One Dimension

nopt

1 a-1

p2

p3

p4

p1

pl = p1 + p2 pr = p3 + p4

a-1

nl nr

n1 n2

pl = p1 pr = p2

Buffer Assignment: Generalized

n3 n4

Indexing 3D objects in Wavelets

A Naïve method: using a 4D R-treePosition of the wavelet coefficient and Magnitude of the coefficientEach vertex needs a number of neighbor vertices tooRetrieval is a two step process: Retrieve those coefficients that fall inside query window Extend the query window to retrieve the neighbors

Our method: indexing the support region

Support Regions of Wavelets

v1v2

v3

v4

v5v6

An Efficient Access Method

w

x

y

Query : ( R, 1.0, 0.7 )

Query : ( R, 0.7, 0.0 )

w = 1.0

w = 0

Experimental Setup

A city is augmented by complex 3D objects such as spheres, pyramids.

Three-dimensional objects are decomposed using wavelet-based techniques and stored in a server.

Clients make a tour in the city from a randomly selected source towards a destination: Using a Tram or on Foot

Experiment Parameters

Data size: 20MB, 40MB, 60MB, 80MB

Query Frame: 5%, 15%, 15%, 20% in height and width

Wireless bandwidth 256Kbps and latency 200ms

Continuous Retrieval

Effect of speed on data retrieval

Index

Effect of speed

Buffer Management

(a) Cache hit rate (b) Data utilization

Effect of buffer size

Overall System Performance

(a) Tram (b) Walk

Query response time (Uniform)

Overall System Performance

(a) Tram (b) Walk

Query response time (Zipf)

Conclusions & Future Work

We proposed a motion-aware approach to continuous retrieval of 3D objects. Experiments shows that the motion aware techniques outperforms traditional ones and overall we achieve high improvement, especially when the view is moving fast.

Future work:Server-side buffer managementReflecting pathways in indices