+




    A Framework for Crowdsourced
    Multimedia Processing and Querying
    Alessandro Bozzon, Ilio Catallo, Eleonora Ciceri, Piero Fraternali,
    Davide Martinenghi, Marco Tagliasacchi

                                                                          0
+                                        1

    CUbRIK Project

       CUbRIK is a research project
        financed by the European Union

       Goals:
         Advance the architecture of
          multimedia search
         Exploit the human
          contribution in multimedia
          search
         Use open-source
          components provided by the
          community
         Start up a search business
          ecosystem

       http://www.cubrikproject.eu/
+                                                                              2

    Humans in Multimedia Information
    Retrieval
       Problem: the uncertainty of analysis algorithms leads to low
        confidence results and conflicting opinions on automatically
        extracted features

       Solution: humans have superior capacity for understanding the
        content of audiovisual material
           State of the art: humans replace automatic feature extraction
            processes (human annotations)




           Our contribution: integration of human judgment and algorithms
             Goal: improve the performance of multimedia content processing
+ Example of CUbRIK Human-enhanced                                      3



 computation: Trademark Logo Detection

    Problem statement: identifying occurrences of trademark logos in
     a video collection through keyword-based queries
        Special case of the classic problem of object recognition




    Use case: a professional user wants to retrieve all the
     occurrences of logos in a large collection of video clips

    Applications: rating effectiveness of advertising, subliminal
     advertising detection, automatic annotation, trademark violation
     detection
+                                                                                  4

    Trademark Logo Detection: problems in
    automatic logo detection
       Problems in automatic logo detection:
           Object recognition is affected by the quality of the input set of
            images




           Uncertain matches, i.e., the ones with low matching score, could not
            contain the searched logo
+                                                                     5

    Trademark Logo Detection:
    contribution of human computation
       Contribution in human computation
           Filter the input logos, eliminating the irrelevant ones
           Segment the input logos




           Validate the matching results
+                                        6

    Trademark Logo Detection: pipeline
+                                   7

    The CrowdSearch framework for
    HC task management
+                                                                                  8

    CrowdSearch framework in the
    Logo detection application
              Problem solving
                 process
    Process




                Task        Crowd
                             Task
                                    Types of tasks
                                    • Automatic tasks
                                    • Crowd tasks: tasks that are executed by an
                                       open-ended community of performers
               Crowd Task
+                                                                                    9

    Community of Performers
     Content edges,
     e.g., IS-A, part.of   Content elements
                                                  The application is deployed as a
                                                  Facebook application

                                                  Seed community
                                                  Information Technology
                           Performer to content   department of Politecnico di
                           edges, e.g., topical
                           group membership
                                                  Milano
          Performers
         edges, e.g.,
           friendship,                            Task propagation
             weak ties
                             Performers           Each user in the seed
                                                  community can propagate
                                                  tasks through the social
                                                  networks
+                                                                              10

    Design of “Validate Logo Images”
                 The “LIKE” task variant requires to choose
                 relevant logos among a set of not filtered images




    Human Task
      Design

                 The “ADD”task variant requires to add new
                 relevant image URLs
                                               Please add new relevant logos
                                                URL…


                                                                  Send
+                                                                                     11

    People to task matching & Task
    Assignment
Task Deployment Criteria         Execution criteria
                                 Constraints of task execution
     Content Affinity Criteria
                                          Time budget for the experiment
       Execution Criteria
                                 Content Affinity criteria
                                 Query on a representation of the users’ capacities
                                 • Current state: manual selection of users
                   People to     • Future work: Geocultural affinity
                 task matching
                                 Questions are dispatched to the crowd according to the
                                 user experience in answering questions
                                 • Expert user: an user that has already answered to
                                   three questions

                       Task      New users answer to “LIKE” questions
                    assignment
                                 Expert users answer to “LIKE”+“ADD” questions
+                                                12

    Task execution
      Task
    execution


      “LIKE” task variant   “ADD” task variant
+                                                              13

    Output aggregation

                                  “LIKE” task variants
                                  Top-5 rated logos are
                                  selected as relevant logos
      Task                        “ADD” task variants
    execution                     New images are fed back to
                                  the LIKE tasks
                Task outputs


                               Task output

                  Output
                aggregation
+                                                             14

    Experimental evaluation

       Three experimental settings:
           No human intervention
           Logo validation performed by two domain experts
           Inclusion of the actual crowd knowledge

       Crowd involvement
           40 people involved
           50 task instances generated
           70 collected answers
+                                                                                                             15

    Experimental evaluation

              1

             0.9

             0.8
                                        Crowd
             0.7
                                                                                 Experts
             0.6
                                                                              Experts
    Recall




                                          Experts
             0.5                                                                                     Aleve
             0.4                    Crowd                                                            Chunky
             0.3
                            No Crowd                                                                 Shout
             0.2                                      Crowd             No Crowd
             0.1

              0        No Crowd
                   0      0.1     0.2     0.3   0.4      0.5      0.6   0.7      0.8       0.9   1

                                                      Precision
+                                                                                                          16

    Experimental evaluation

              1

             0.9

             0.8                                               Precision decreases
                                        Crowd
             0.7
                                                                              Experts
             0.6                                               Reasons for the wrong inclusion
                                                                      Experts
    Recall




                                          Experts              • Geographical location of the users
             0.5                                                                             Aleve
                                                               • Expertise of the involved users
             0.4                    Crowd                                                         Chunky
             0.3
                            No Crowd                                                              Shout
             0.2                                      Crowd             No Crowd
             0.1

              0        No Crowd
                   0      0.1     0.2     0.3   0.4      0.5      0.6   0.7   0.8       0.9   1

                                                      Precision
+                                                                                                             17

    Experimental evaluation

              1
                                                                        Precision decreases
                                                                        • Similarity between two
             0.9
                                                                           logos in the data set
             0.8
                                        Crowd
             0.7
                                                                                 Experts
             0.6
                                                                              Experts
    Recall




                                          Experts
             0.5                                                                                     Aleve
             0.4                    Crowd                                                            Chunky
             0.3
                            No Crowd                                                                 Shout
             0.2                                      Crowd             No Crowd
             0.1

              0        No Crowd
                   0      0.1     0.2     0.3   0.4      0.5      0.6   0.7      0.8       0.9   1

                                                      Precision
+                                                                                                  18

    Future directions

       Task design:
           Implement new task types (tag / comment / like / add / modify…)
           Partition large task instances into several smaller instances dispatched to multiple
            users

       Task assignment: study how to associate the most suitable request with
        the most appropriate user
           Implement a ranking function on worker pool, based on the
            expertise, geocultural information and past work history of the performers

       Task execution: multiple heterogeneous platforms
        (Facebook, LinkedIn, Twitter, stand-alone application)

       More use cases:
           Breaking news
           Fashion trend

The CrowdSearch framework

  • 1.
    + A Framework for Crowdsourced Multimedia Processing and Querying Alessandro Bozzon, Ilio Catallo, Eleonora Ciceri, Piero Fraternali, Davide Martinenghi, Marco Tagliasacchi 0
  • 2.
    + 1 CUbRIK Project  CUbRIK is a research project financed by the European Union  Goals:  Advance the architecture of multimedia search  Exploit the human contribution in multimedia search  Use open-source components provided by the community  Start up a search business ecosystem  http://www.cubrikproject.eu/
  • 3.
    + 2 Humans in Multimedia Information Retrieval  Problem: the uncertainty of analysis algorithms leads to low confidence results and conflicting opinions on automatically extracted features  Solution: humans have superior capacity for understanding the content of audiovisual material  State of the art: humans replace automatic feature extraction processes (human annotations)  Our contribution: integration of human judgment and algorithms  Goal: improve the performance of multimedia content processing
  • 4.
    + Example ofCUbRIK Human-enhanced 3 computation: Trademark Logo Detection  Problem statement: identifying occurrences of trademark logos in a video collection through keyword-based queries  Special case of the classic problem of object recognition  Use case: a professional user wants to retrieve all the occurrences of logos in a large collection of video clips  Applications: rating effectiveness of advertising, subliminal advertising detection, automatic annotation, trademark violation detection
  • 5.
    + 4 Trademark Logo Detection: problems in automatic logo detection  Problems in automatic logo detection:  Object recognition is affected by the quality of the input set of images  Uncertain matches, i.e., the ones with low matching score, could not contain the searched logo
  • 6.
    + 5 Trademark Logo Detection: contribution of human computation  Contribution in human computation  Filter the input logos, eliminating the irrelevant ones  Segment the input logos  Validate the matching results
  • 7.
    + 6 Trademark Logo Detection: pipeline
  • 8.
    + 7 The CrowdSearch framework for HC task management
  • 9.
    + 8 CrowdSearch framework in the Logo detection application Problem solving process Process Task Crowd Task Types of tasks • Automatic tasks • Crowd tasks: tasks that are executed by an open-ended community of performers Crowd Task
  • 10.
    + 9 Community of Performers Content edges, e.g., IS-A, part.of Content elements The application is deployed as a Facebook application Seed community Information Technology Performer to content department of Politecnico di edges, e.g., topical group membership Milano Performers edges, e.g., friendship, Task propagation weak ties Performers Each user in the seed community can propagate tasks through the social networks
  • 11.
    + 10 Design of “Validate Logo Images” The “LIKE” task variant requires to choose relevant logos among a set of not filtered images Human Task Design The “ADD”task variant requires to add new relevant image URLs Please add new relevant logos URL… Send
  • 12.
    + 11 People to task matching & Task Assignment Task Deployment Criteria Execution criteria Constraints of task execution Content Affinity Criteria Time budget for the experiment Execution Criteria Content Affinity criteria Query on a representation of the users’ capacities • Current state: manual selection of users People to • Future work: Geocultural affinity task matching Questions are dispatched to the crowd according to the user experience in answering questions • Expert user: an user that has already answered to three questions Task New users answer to “LIKE” questions assignment Expert users answer to “LIKE”+“ADD” questions
  • 13.
    + 12 Task execution Task execution “LIKE” task variant “ADD” task variant
  • 14.
    + 13 Output aggregation “LIKE” task variants Top-5 rated logos are selected as relevant logos Task “ADD” task variants execution New images are fed back to the LIKE tasks Task outputs Task output Output aggregation
  • 15.
    + 14 Experimental evaluation  Three experimental settings:  No human intervention  Logo validation performed by two domain experts  Inclusion of the actual crowd knowledge  Crowd involvement  40 people involved  50 task instances generated  70 collected answers
  • 16.
    + 15 Experimental evaluation 1 0.9 0.8 Crowd 0.7 Experts 0.6 Experts Recall Experts 0.5 Aleve 0.4 Crowd Chunky 0.3 No Crowd Shout 0.2 Crowd No Crowd 0.1 0 No Crowd 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Precision
  • 17.
    + 16 Experimental evaluation 1 0.9 0.8 Precision decreases Crowd 0.7 Experts 0.6 Reasons for the wrong inclusion Experts Recall Experts • Geographical location of the users 0.5 Aleve • Expertise of the involved users 0.4 Crowd Chunky 0.3 No Crowd Shout 0.2 Crowd No Crowd 0.1 0 No Crowd 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Precision
  • 18.
    + 17 Experimental evaluation 1 Precision decreases • Similarity between two 0.9 logos in the data set 0.8 Crowd 0.7 Experts 0.6 Experts Recall Experts 0.5 Aleve 0.4 Crowd Chunky 0.3 No Crowd Shout 0.2 Crowd No Crowd 0.1 0 No Crowd 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Precision
  • 19.
    + 18 Future directions  Task design:  Implement new task types (tag / comment / like / add / modify…)  Partition large task instances into several smaller instances dispatched to multiple users  Task assignment: study how to associate the most suitable request with the most appropriate user  Implement a ranking function on worker pool, based on the expertise, geocultural information and past work history of the performers  Task execution: multiple heterogeneous platforms (Facebook, LinkedIn, Twitter, stand-alone application)  More use cases:  Breaking news  Fashion trend