Enhancing Online 3D Products through Crowdsourcing Thi Phuong - - PowerPoint PPT Presentation

▶
enhancing online 3d products through crowdsourcing
SMART_READER_LITE
LIVE PREVIEW

Enhancing Online 3D Products through Crowdsourcing Thi Phuong - - PowerPoint PPT Presentation

Oct 24, 2022 108 likes •571 views

Enhancing Online 3D Products through Crowdsourcing Thi Phuong Nghiem, Axel Carlier, Geraldine Morin Vincent Charvillat University of Toulouse IRIT ENSEEIHT ACM Workshop on Crowdsourcing for mmedia - ACM MM'12 ACM Workshop on Crowdsourcing for

slide-1
SLIDE 1

Enhancing Online 3D Products through Crowdsourcing

Thi Phuong Nghiem, Axel Carlier, Geraldine Morin Vincent Charvillat

University of Toulouse IRIT ENSEEIHT

ACM Workshop on Crowdsourcing for mmedia - ACM MM'12 ACM Workshop on Crowdsourcing for mmedia - ACM MM'12

slide-2
SLIDE 2

2

Outline

  • Motivation through three initial observations
  • Our crowdsourcing set-up for 3D content
  • Results
  • Conclusion and perspective
slide-3
SLIDE 3

Observation #1

  • Crowdsourcing definition
  • Workers may be motivated by ...
  • Entertainment/personal enjoyment
  • Altruism
  • Financial reward
  • Specific incentives in an E-commerce set-up ?
  • Would you like free PnP ?
  • Discount code ?
slide-4
SLIDE 4

Observation #2

  • E-commerce Tasks: what could be outsourced to customers ?
  • [Little et al. 10] proposed two main categories :
  • Decision tasks (opinion production, rating, comparison)
  • Creation Tasks (content production, composition, edition)
slide-5
SLIDE 5

Observation #2 (cont.)

slide-6
SLIDE 6

Observation #2 (cont.)

slide-7
SLIDE 7

Observation #3(D)

  • 3D content and E-commerce: an emerging topic ?
  • 3D product presentation as an alternative to image and/or video
slide-8
SLIDE 8

Observation #3(D)

  • 3D content and E-commerce: an emerging topic ?
  • 3D product presentation as an alternative to image and/or video
slide-9
SLIDE 9

Observation #3(D)

  • 3D content and E-commerce: an emerging topic ?
  • 3D product presentation as an alternative to image and/or video
slide-10
SLIDE 10

10

Our idea

Product Features Textual Description Product Features Textual Description 3D Model

  • f the Product

3D Model

  • f the Product

Customers Motivated workers Customers Motivated workers Product Features Description Product Features Description 3D model

  • f a product

3D model

  • f a product

Semantic links Semantic links Preferences Preferences Representative Views Representative Views

slide-11
SLIDE 11

11

A product webpage with text and 3D

Product Features Textual Description Product Features Textual Description 3D Model

  • f the Product

3D Model

  • f the Product

Customers Motivated workers Customers Motivated workers Product Features Description Product Features Description 3D model

  • f a product

3D model

  • f a product

Semantic links Semantic links Preferences Preferences Representative Views Representative Views

We use x3dom to render 3D objects (plug-in free solution).

slide-12
SLIDE 12

12

Linking product features and representative 3D views

(1) Crowdsourced tasks: Locate each visible feature on the 3D model (2) (1) (2) Use crowdsourced data to compute Recommended Views and create Semantic Link Semantic Links are presented as blue bullets with question mark ”?” These links help gathering knowledge about a product. They also ease browsing its 3D model.

slide-13
SLIDE 13

13

Semantic Link Definition

  • A Semantic Link matches a product feature description with its

possibly visible position on the 3D model.

  • Example of Semantic Link for Mode Dial feature

Semantic Link

slide-14
SLIDE 14

14

Basic Task

(1) A product feature (text item) is selected using checkbox. (2) 3D objects are manipulated with mouse interactions (zoom/pan/rotate). (2) Feature position is located by double-clicking on it and a red dot appears to mark the position (called marked-point) on the 3D object (3) Level of expertise/confidence is indicated by the worker

slide-15
SLIDE 15

15

Video Part #1

slide-16
SLIDE 16
slide-17
SLIDE 17

17

Crowdsourced Data

  • Collected data are as follows:
  • The selected (textual) feature
  • The amount of time it took the user to find each feature's

visualization in 3D.

  • The world coordinates of the 3D marked-point, of its normal

vector and its camera position (if done or "I don't know" events if users cannot locate the feature)

  • Level of expertise about each product.
  • Types of events created by users (zoom/pan/rotate/click).
slide-18
SLIDE 18

18

Recommended View Computation

  • We aim at producing one representative view for each visible feature
  • To generate a Recommended View, we compute the camera

parameters so as its look-at point is the median of marked-points from the 'crowd'.

  • The quality of recommended views is correlated with the dispersion
  • f marked-points.

Recommended view for the 'mode button'

slide-19
SLIDE 19

19

?

slide-20
SLIDE 20

20

The 4 Steps

(1) Association / Linking : what we just saw ! (2) Exploitation / Evaluation Same operations as in (1), except that each time users choose a feature, a recommended view is automatically displayed to suggest the corresponding 3D visualization of the feature. Then, for each feature, we ask the user if the recommended view was helpful or not. (3) Helpfullness evaluation Same as (2) except that they are given the results of the helpfulness evaluation of the recommended views from Part 2. (4) New interface evaluation We assess that the enriched interface is better than the initial one (3D product along with its textual description)

slide-21
SLIDE 21

21

The 4 Steps

(1) Association / Linking : what we just saw ! (2) Exploitation / Evaluation Same operations as in (1), except that each time users choose a feature, a recommended view is automatically displayed to suggest the corresponding 3D visualization of the feature. Then, for each feature, we ask the user if the recommended view was helpful or not. (3) Helpfullness evaluation Same as (2) except that they are given the results of the helpfulness evaluation of the recommended views from Part 2. (4) New interface evaluation We assess that the enriched interface is better than the initial one (3D product along with its textual description)

slide-22
SLIDE 22

22

Recommended View Integration - Part (2)

Recommendation is given automatically to user at each textual name selection

slide-23
SLIDE 23

23

Video Part #2

slide-24
SLIDE 24

24

The 4 Steps

(1) Association / Linking : what we just saw ! (2) Exploitation / Evaluation Same operations as in (1), except that each time users choose a feature, a recommended view is automatically displayed to suggest the corresponding 3D visualization of the feature. Then, for each feature, we ask the user if the recommended view was helpful or not. (3) Helpfullness evaluation Same as (2) except that they are given the results of the helpfulness evaluation of the recommended views from Part 2. (4) New interface evaluation We assess that the enriched interface is better than the initial one (3D product along with its textual description)

slide-25
SLIDE 25

25

Helpfulness Influence (Part 3)

Recommended View is integrated with Helpfulness Score at each textual name selection

slide-26
SLIDE 26

26

Video Part #3

slide-27
SLIDE 27

27

The 4 Steps

(1) Association / Linking : what we just saw ! (2) Exploitation / Evaluation Same operations as in (1), except that each time users choose a feature, a recommended view is automatically displayed to suggest the corresponding 3D visualization of the feature. Then, for each feature, we ask the user if the recommended view was helpful or not. (3) Helpfullness evaluation Same as (2) except that they are given the results of the helpfulness evaluation of the recommended views from Part 2. (4) New interface evaluation We assess that the enriched interface is better than the initial one (3D product along with its textual description)

slide-28
SLIDE 28

28

Semantic Link Creation (Part 4)

Recommended View for Mode Dial is shown when its Semantic Link is clicked

slide-29
SLIDE 29

29

Video Part #4

slide-30
SLIDE 30

30

Protocol of User Study

  • Protocol: 82 participants (47 males + 35 females, [19-40 y.] ) divided

into 4 parts:

  • Outsourcing tasks are first given to users (Part 1 - 20).
  • Outsourcing tasks with recommendations (Part 2 - 28).

– All types of users – Expert users

  • Helpfulness evaluation of Recommended Views (Part 3 - 14).
  • Novel interface evaluation (Part 4 - 20).
slide-31
SLIDE 31

31

Our Models

  • Six models from evermotion (see http://www.evermotion.org)
  • Different complexity,

aesthetics

  • Easy/ Simple /common

(coffee m. filter holder)

  • Technical details

electric guitars (jack socket)

  • Stabilizer switch of the

camera is hard to find

slide-32
SLIDE 32

32

Results

slide-33
SLIDE 33

33

Results

  • Average time for users to locate a feature in 3D:

The recommendation helps users to execute the tasks quicker (part 2).

slide-34
SLIDE 34

34

Results

  • Recommendation helps users to execute the tasks more efficient:

The number of good answers increases while the number of wrong answers decrease (part 2).

slide-35
SLIDE 35

35

Results

  • Influence of helpfulness evaluation (part 3):

Part 3 gets less ”I don't know” answers for Technical Features (eg. jack socket-guitar) and less ”Wrong” answers for Hard Features (eg. stabilizer switch) than part 2 does.

slide-36
SLIDE 36

36

Novel Interface Evaluation

  • Percentage of users who prefer our enriched interface (part 4):
slide-37
SLIDE 37

37

Conclusion Our experiments have shown two particular points:

  • first, the proposed enhancement of the 3D

interactive product has proved useful in two ways:

  • qualitatively, users have appreciated it,
  • quantitatively, it has improved their

performances.

  • Second, crowdsourcing has proved useful in this

context …

slide-38
SLIDE 38

38

Could this semantic association be created by the seller or an expert hired by the seller ?

  • Does it scale ? 15-20 features per product,

thousands of product …

  • The experts are expensive (and may be biased

towards advertising the products).

  • Our experts are inexpensive and the

customers are the best possible "interesting features detectors".

slide-39
SLIDE 39

39

Perspectives

  • Deepen analysis on the crowd's input:

See the influence of initial viewpoints. See the conditions so as the workers give up or submit a wrong answer.

  • Outsource more tasks:

Rank the features by order of importance & Create an automatic description of a 3D product that would emphasize on the more popular features

slide-40
SLIDE 40

40

Towards Smart Products/Graphics

slide-41
SLIDE 41

41

Thank you for your attention

slide-42
SLIDE 42

42

Crowdsourcing Tasks (Part 1)

20 participants are asked to select a textual name/description and locate it on the 3D model.

slide-43
SLIDE 43

43

Helpfulness Evaluation (Part 2)

Users are asked to evaluate the helpfulness of Recommended View at the end of each task

slide-44
SLIDE 44

A bit of context

Consider a challenging visual task:

  • bject detection

where are we now ?

slide-45
SLIDE 45

Pascal VOC (Visual Obj. Classes)

  • The goal of this challenge is to recognize objects from a

number of visual object classes in realistic scenes (i.e.not presegmented objects).

  • The four classes in 2005: motorbikes, bicycles, people, cars
  • Two main competitions: presence/absence AND detection
  • Only 1578 images
  • Twenty classes since 2007

The PASCAL Visual Object Classes (VOC) Challenge International Journal of Computer Vision, 88(2), 303-338, 2010 The PASCAL Visual Object Classes (VOC) Challenge International Journal of Computer Vision, 88(2), 303-338, 2010

slide-46
SLIDE 46

Pascal VOC (Visual Obj. Classes)

  • The goal of this challenge is to recognize objects from a

number of visual object classes in realistic scenes (i.e.not presegmented objects).

  • The four classes in 2005: motorbikes, bicycles, people, cars
  • Two main competitions: presence/absence AND detection
  • Only 1578 images
  • Twenty classes since 2007

The PASCAL Visual Object Classes (VOC) Challenge International Journal of Computer Vision, 88(2), 303-338, 2010 The PASCAL Visual Object Classes (VOC) Challenge International Journal of Computer Vision, 88(2), 303-338, 2010