Predictive Video Retrieval A Matter of Trust Bouke Huurnink - - PowerPoint PPT Presentation

predictive video retrieval
SMART_READER_LITE
LIVE PREVIEW

Predictive Video Retrieval A Matter of Trust Bouke Huurnink - - PowerPoint PPT Presentation

Predictive Video Retrieval A Matter of Trust Bouke Huurnink MediaMill The Team Bouke Huurnink Michiel van Liempt Jiyin He Richard van Balen Koen van de Sande FeiYan Ork de Rooij Muhammad Tahir Cees Snoek Krystian Mikolajczyk Maarten


slide-1
SLIDE 1

Predictive Video Retrieval

A Matter of Trust

Bouke Huurnink

MediaMill

slide-2
SLIDE 2

The Team

Bouke Huurnink Jiyin He Koen van de Sande Ork de Rooij Cees Snoek Maarten de Rijke Jan van Gemert Jasper Uijlings Xirong Li Ivo Everts Vladimir Nedovic Michiel van Liempt Richard van Balen FeiYan Muhammad Tahir Krystian Mikolajczyk Josef Kittler Jan-Mark Geusebroek Theo Gevers Marcel Worring Arnold Smeulders Dennis Koelma

slide-3
SLIDE 3

Come see our interactive search demo

0.05 0.10 0.15 0.20 0.05 0.10 0.15 0.20

Now with (inter)active learning! Presented by Ork de Rooij

UvA

slide-4
SLIDE 4

Why predictive video retrieval?

  • Video retrieval is multichannel problem:
  • Speech
  • Detectors
  • Examples
  • Observations
  • Speech works for named entity topics
  • Detectors work when closely related to topic
  • Examples can also work pretty well
  • We want to exploit this knowledge
slide-5
SLIDE 5

Idea

  • Predict which type of search - retrieval

channel - we can trust for a topic

  • Rerank results from this channel with

secondary result information

slide-6
SLIDE 6

Outline

  • System description
  • Result overview
  • Analysis
  • Conclusion
slide-7
SLIDE 7

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results
slide-8
SLIDE 8

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results
slide-9
SLIDE 9

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results

Distribute ASR and MT over shot neighbourhood, then retrieval using language modelling approach Pseudo active-learning, with positive examples from topic and 100 random negative examples from collection Content based selection from 57 learned concepts, followed by unweighted score-based fusion

slide-10
SLIDE 10

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results
slide-11
SLIDE 11

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results

Named entity? Trust speech results Detector match? Trust detector results Else...trust example results

slide-12
SLIDE 12

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results
slide-13
SLIDE 13

Our predictive system

Retrieval Channels Speech Search Detector Search Example Search Predict Trusted Channel Reranking Final Results Information Need

Find shots of pieces

  • f paper with writing,

typing, or printing, filling more than half

  • f the frame area.

Result Lists Trusted Results

  • Detector results

Secondary Results

  • Speech results

Secondary Results

  • Example results

Truncate result lists to top 1000 Eliminate all results not in trusted list Combine results with (weighted) Borda fusion

slide-14
SLIDE 14

Query-class vs Prediction

Query-class Prediction

Query class determines retrieval strategy Query features determine retrieval strategy Focus on assigning query- class dependent weights Focus on identifying trusted retrieval channel

slide-15
SLIDE 15

Runs

  • Speech channel only UvA-MM-6
  • Detector channel only UvA-MM-5
  • Example channel only supplementary
  • Predictive reranking UvA-MM-4
  • Predictive weighted reranking UvA-MM-3
slide-16
SLIDE 16

0.01 0.02 0.03 0.04 0.05 0.06 0.07

Overall Automatic Search Performance

Predictive reranking Detector channel Example channel Speech channel Predictive weighted reranking

mean inferred average precision All runs

slide-17
SLIDE 17

0.01 0.02 0.03 0.04 0.05 0.06 0.07

Overall Automatic Search Performance

Predictive reranking Detector channel Example channel Speech channel Predictive weighted reranking

Predictive reranking

  • utperforms individual channels

mean inferred average precision All runs

slide-18
SLIDE 18

0.01 0.02 0.03 0.04 0.05 0.06 0.07

Overall Automatic Search Performance

Predictive reranking Detector channel Example channel Speech channel Predictive weighted reranking

Predictive reranking

  • utperforms individual channels

Weighting did not have big influence

mean inferred average precision All runs

slide-19
SLIDE 19

General findings

  • 20 topics > 0.05 inferred average precision
  • 1 speech topic
  • 11 detector topics
  • 8 example topics
  • Accurately predicted 15 of 20 topics
slide-20
SLIDE 20

A closer look

person opening door a bridge people with trees and plants face filling over half the frame paper with writing people with a body of water a map vehicle moving away people looking in microscope person watching television people in a kitchen a crowd of people outdoors a classroom scene an airplane exterior a plant that is the main object a street scene at night people at table with computer people in white lab coats ships or boats in the water man talking to camera indoors

inferred average precision

0.1 0.2 0.3 0.4 0.5

Predictive w. reranking Detector channel Example channel Speech channel

slide-21
SLIDE 21

A closer look

person opening door a bridge people with trees and plants face filling over half the frame paper with writing people with a body of water a map vehicle moving away people looking in microscope person watching television people in a kitchen a crowd of people outdoors a classroom scene an airplane exterior a plant that is the main object a street scene at night people at table with computer people in white lab coats ships or boats in the water man talking to camera indoors

A lot of variance between channels

inferred average precision

0.1 0.2 0.3 0.4 0.5

Predictive w. reranking Detector channel Example channel Speech channel

slide-22
SLIDE 22

person opening door a bridge people with trees and plants paper with writing a map people looking in microscope people in a kitchen a crowd of people outdoors a classroom scene an airplane exterior a plant that is the main object a street scene at night people at table with computer people in white lab coats man talking to camera indoors

When prediction worked

Only trusted channel and reranked performance shown

Predictive w. reranking Detector channel Example channel Speech channel inferred average precision

0.1 0.2 0.3 0.4 0.5

slide-23
SLIDE 23

person opening door a bridge people with trees and plants paper with writing a map people looking in microscope people in a kitchen a crowd of people outdoors a classroom scene an airplane exterior a plant that is the main object a street scene at night people at table with computer people in white lab coats man talking to camera indoors

When prediction worked

Only trusted channel and reranked performance shown

Predictive reranking often close to or better than trusted channel

Predictive w. reranking Detector channel Example channel Speech channel inferred average precision

0.1 0.2 0.3 0.4 0.5

slide-24
SLIDE 24

When prediction didn’t work

face filling over half the frame people with a body of water vehicle moving away person watching television ships or boats in the water man talking to camera indoors

Predictive w. reranking Detector channel Example channel Speech channel inferred average precision

0.1 0.2 0.3 0.4 0.5

Only trusted channel and reranked performance shown

slide-25
SLIDE 25

When prediction didn’t work

face filling over half the frame people with a body of water vehicle moving away person watching television ships or boats in the water man talking to camera indoors

Predictive w. reranking Detector channel Example channel Speech channel

Predictive reranking boosts trusted channel results

inferred average precision

0.1 0.2 0.3 0.4 0.5

Only trusted channel and reranked performance shown

slide-26
SLIDE 26

Conclusions

slide-27
SLIDE 27

Conclusions

Predictive retrieval works, even with simple reranking

slide-28
SLIDE 28

Conclusions

Predictive retrieval works, even with simple reranking Incorrect predictions have limited impact

slide-29
SLIDE 29

Conclusions

Predictive retrieval works, even with simple reranking Incorrect predictions have limited impact Good ingredients are crucial: garbage in garbage out!

slide-30
SLIDE 30

Beeld en Geluid Searches

20 20 uur 20 uur journaal aartsen afghanistan ajax algemene beschouwingen amsterdam

andere tijden avondjournaal balkenende beatrix buitenhof bush close up de wereld draait door debat eenvandaag evn feyenoord gemeenteraadsverkiezingen

goedemorgen nederland hirsi ali holland sport holleeder internationale nieuwsuitwisseling irak

iran jeugdjournaal journaal journaal 20 kassa klokhuis

koefnoen koninginnedag kooten kopspijkers kro kruispunt langs de lijn libanon lijst 0 lingo man bijt hond

max catherine maxima mens milosevic miniatuur moszkowicz nederland kiest netwerk nieuwslicht nioscoop nos nos journaal nova nps arena opsporing verzocht paul de leeuw pauw pauw en

witteman pauw witteman pechtold politie polygoon radar rembrandt rouvoet rutte saddam schepper

co schepper en co schipholbrand sesamstraat sonja spiritus sporen uit het oosten sport sportjournaal

studio sport tegenlicht televisie tros tv show twee vandaag uruzgan vandaag

verdonk verkiezingen voetbal vragenuur vragenuurtje vroege vogels wereld draait door wilders wouter bos

zembla zoekt en gij zult vinden zomergasten

Pondering

  • What if we had more variety in query types?

General

  • bject queries

Named entity queries