SlideShare ist ein Scribd-Unternehmen logo
1 von 13
Downloaden Sie, um offline zu lesen
Towards Recommender Systems
for Police Photo Lineup
Ladislav Peškaa and Hana Trojanováb
a Department of Software Engineering,
b Department of Psychology,
Charles University, Prague,
Czech Republic
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
2
Recommender Systems
 Propose relevant items to the right persons at the right time
 Successful on many various domains
 Multimedia
 E-commerce
 POIs
 Events
 Recepies
 …
 However, some novel (relevant, important) tasks are yet to
discover
DLRS@RecSys2017, Como
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
3
Police Photo Lineup
 Eyewitness identification of the suspect / offender during
criminal proceedings
 Most common case: select suspect among other persons (fillers)
 Either „in natura“ or based on photographies
DLRS@RecSys2017, Como
- Suspect is positioned at
random
- 3-7 fillers are added, witness
should not know them
- Wittness is instructed that the
offender may or may not be
present
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
4
Errors in Photo Lineup
DLRS@RecSys2017, Como
1) Observe the event
2) Identify the offender
- among other candidates
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
5
Errors in Photo Lineup
DLRS@RecSys2017, Como
1) Observe the event.
- numerous sources of error (poor lighting,
large distance, short time, fear, fatigue…)
- affects memory and recognition processes
- decrease witness certainty and reliability
2) Identify the offender
- decreased certainty may lead
to incorrect identification
- and conviction of innocent
persons
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
6
Assembling Fair Photo Lineups
 System should be set to eliminate testimony of uncertain
witnesses
 Double-blind administration
 Fair (unbiased) assembling of lineups
 Out of the dataset of candidates, lineup administrator should select
fillers similar to the suspect
 Currently, only feature-based filtering systems are available
 Better automatization in content processing is necessary
 How to better approximate human-perceived similarity of
persons?
 Convolutional Neural Networks?
DLRS@RecSys2017, Como
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
7
Assembling Fair Photo Lineups
 First approach: item-based recommendations
 Propose top-k candidates for each suspect
 CB-RS, cosine similarity of candidates’ and suspect’s CB features
 Nationality, Age, Appearence features
 Baseline (mimic feature-based filters)
 Visual-RS, cosine similarity of candidates’ and suspect’s VGG-
Face probability layers
 Approximate the similarity of a candidate and the suspect through their
similarities with other persons
 VGG-Face1 network
 VGG network for facial recognition problem
 Trained as a multi-classification problem
 Dataset of 2,622 celebrities with 1,000 images each
1Parkhi, O. M., Vedaldi, A., & Zisserman, A. (2015). Deep Face Recognition. British Machine Vision Conference.
DLRS@RecSys2017, Como
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
8
Research Questions
 Is candidates recommendation (based on VGG-Face’s
features) suitable for the lineup assembling?
 How well is the human-perceived similarity approximated within top-k
candidates?
 What is the role of diversity in lineup assembling?
 Is there a space for personalized recommendations (i.e. personalized per
lineup administrator)?
 Can there be „too similar“ candidates?
DLRS@RecSys2017, Como
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
9
Evaluation – Protocol
 Dataset of missing and wanted persons in Czech Republic
 4423 males
 User study, 30 different lineups, 7 forensic technicians
 For each lineup, both CB-RS and Visual-RS proposed top-20 candidates
 Merged into one list, forensic technician should select relevant fillers for
particular suspect
 Follow-up questionaire and discussion
DLRS@RecSys2017, Como
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
10
Results
 In total 202 completed lineups, 800 selected candidates
 Recommendations were seemingly relevant in most cases
 Visual-RS considerably outperformed CB-RS
 The performance difference was much smaller for suspects from outside
of central Europe
 Effect of different ethnicity? (Known in the forensic psychology literature).
DLRS@RecSys2017, Como
Total lineups Selected candidates Level of agreement
All lineups
Visual-RS
202
466 (58%) 0.178
CB-RS 298 (37%) 0.138
Lineups with suspects from countries outside Central Europe
Visual-RS
53
105 (51%) 0.128
CB-RS 82 (40%) 0.205
Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
11
Results II.
 Very low intersection (1.8%) of top-20 CB-RS and Visual-RS candidates
 Potential to apply some combination in the future work
 Average position of selected candidate
CB-RS: 8.9 (+/- 5.2); Visual-RS: 8.2 (+/- 5.7)
 Ordering seems relevant even in top-20, however can be improved
 Low level of agreement among participants on the selected candidates (0.178, 0.138)
 Limits of the global ordering?
 Potential for personalized recommendation?
 Less diversity is better
 Participants agreed on the need for providing homogeneous lineups
 Dynamical recommendation based on the selected candidates?
 No seemingly too similar candidates
 However, the dataset contained some duplicates
 Should be addressed in the future work
DLRS@RecSys2017, Como
Conclusions
 Recommending candidates based on CNN’s features seems
plausible in police photo lineup
 Item-based recommendations seems like a good first approach
 There is a potential for:
 Session-based personalized recommendations (need for uniformity of the final list)
 Long-term preference learning (due to the level of disaggreeement among lineup
administrators)
 Approaches combining CNN’s features with content-based attributes
Future work
 Move from recommended candidates to recommended lineups
 Evaluation through mock witnesses
 Deployable demo
DLRS@RecSys2017, Como Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
12
DLRS@RecSys2017, Como Ladislav Peska & Hana Trojanova: Towards
Recommender Systems for Police Photo Lineup
13
Thank you!
Questions, comments?
Supplementary materials: http://github.com/lpeska/lineups, https://tinyurl.com/yat3rtr2
Slides: https://www.slideshare.net/LadislavPeska/towards-recommender-systems-for-police-photo-lineup

Weitere ähnliche Inhalte

Ähnlich wie Towards Recommender Systems for Police Photo Lineup

InstructionsW4 Nightingale Case A & B – 35 points - Individual A.docx
InstructionsW4 Nightingale Case A & B – 35 points - Individual A.docxInstructionsW4 Nightingale Case A & B – 35 points - Individual A.docx
InstructionsW4 Nightingale Case A & B – 35 points - Individual A.docxdirkrplav
 
Dundee Police and Criminal Justice Group Presentation
Dundee Police and Criminal Justice Group PresentationDundee Police and Criminal Justice Group Presentation
Dundee Police and Criminal Justice Group PresentationEric Halford PhD(can)
 
Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)Rich Heimann
 
Data Science and Analytics Brown Bag
Data Science and Analytics Brown BagData Science and Analytics Brown Bag
Data Science and Analytics Brown BagDataTactics
 
Book Reference Peak, K. J. (2015). Policing America C.docx
Book Reference Peak, K. J. (2015). Policing America C.docxBook Reference Peak, K. J. (2015). Policing America C.docx
Book Reference Peak, K. J. (2015). Policing America C.docxAASTHA76
 
Crime Pattern Detection using K-Means Clustering
Crime Pattern Detection using K-Means ClusteringCrime Pattern Detection using K-Means Clustering
Crime Pattern Detection using K-Means ClusteringReuben George
 
Propose Data Mining AR-GA Model to Advance Crime analysis
Propose Data Mining AR-GA Model to Advance Crime analysisPropose Data Mining AR-GA Model to Advance Crime analysis
Propose Data Mining AR-GA Model to Advance Crime analysisIOSR Journals
 

Ähnlich wie Towards Recommender Systems for Police Photo Lineup (11)

InstructionsW4 Nightingale Case A & B – 35 points - Individual A.docx
InstructionsW4 Nightingale Case A & B – 35 points - Individual A.docxInstructionsW4 Nightingale Case A & B – 35 points - Individual A.docx
InstructionsW4 Nightingale Case A & B – 35 points - Individual A.docx
 
Dundee Police and Criminal Justice Group Presentation
Dundee Police and Criminal Justice Group PresentationDundee Police and Criminal Justice Group Presentation
Dundee Police and Criminal Justice Group Presentation
 
Super-recogniser Paper.PDF
Super-recogniser Paper.PDFSuper-recogniser Paper.PDF
Super-recogniser Paper.PDF
 
Dna Evidence Essay
Dna Evidence EssayDna Evidence Essay
Dna Evidence Essay
 
Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)
 
Data Science and Analytics Brown Bag
Data Science and Analytics Brown BagData Science and Analytics Brown Bag
Data Science and Analytics Brown Bag
 
Crime
CrimeCrime
Crime
 
Book Reference Peak, K. J. (2015). Policing America C.docx
Book Reference Peak, K. J. (2015). Policing America C.docxBook Reference Peak, K. J. (2015). Policing America C.docx
Book Reference Peak, K. J. (2015). Policing America C.docx
 
Crime Pattern Detection using K-Means Clustering
Crime Pattern Detection using K-Means ClusteringCrime Pattern Detection using K-Means Clustering
Crime Pattern Detection using K-Means Clustering
 
Propose Data Mining AR-GA Model to Advance Crime analysis
Propose Data Mining AR-GA Model to Advance Crime analysisPropose Data Mining AR-GA Model to Advance Crime analysis
Propose Data Mining AR-GA Model to Advance Crime analysis
 
DNA Forensic Study
DNA Forensic StudyDNA Forensic Study
DNA Forensic Study
 

Mehr von Ladislav Peska

Fuzzy D’Hondt’s Algorithm for On-line Recommendations Aggregation
Fuzzy D’Hondt’s Algorithm for On-line Recommendations AggregationFuzzy D’Hondt’s Algorithm for On-line Recommendations Aggregation
Fuzzy D’Hondt’s Algorithm for On-line Recommendations AggregationLadislav Peska
 
Off-line vs. On-line Evaluation of Recommender Systems in Small E-commerce
Off-line vs. On-line Evaluation of Recommender Systems in Small E-commerceOff-line vs. On-line Evaluation of Recommender Systems in Small E-commerce
Off-line vs. On-line Evaluation of Recommender Systems in Small E-commerceLadislav Peska
 
Linking Content Information with Bayesian Personalized Ranking via Multiple C...
Linking Content Information with Bayesian Personalized Ranking via Multiple C...Linking Content Information with Bayesian Personalized Ranking via Multiple C...
Linking Content Information with Bayesian Personalized Ranking via Multiple C...Ladislav Peska
 
Towards Complex User Feedback and Presentation Context in Recommender Systems
Towards Complex User Feedback and Presentation Context in Recommender SystemsTowards Complex User Feedback and Presentation Context in Recommender Systems
Towards Complex User Feedback and Presentation Context in Recommender SystemsLadislav Peska
 
Using the Context of User Feedback in Recommender Systems
Using the Context of User Feedback in Recommender SystemsUsing the Context of User Feedback in Recommender Systems
Using the Context of User Feedback in Recommender SystemsLadislav Peska
 
How to Interpret Implicit User Feedback
How to Interpret Implicit User FeedbackHow to Interpret Implicit User Feedback
How to Interpret Implicit User FeedbackLadislav Peska
 
Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...Ladislav Peska
 
RecSys Challenge 2014, SemWexMFF group
RecSys Challenge 2014, SemWexMFF groupRecSys Challenge 2014, SemWexMFF group
RecSys Challenge 2014, SemWexMFF groupLadislav Peska
 

Mehr von Ladislav Peska (8)

Fuzzy D’Hondt’s Algorithm for On-line Recommendations Aggregation
Fuzzy D’Hondt’s Algorithm for On-line Recommendations AggregationFuzzy D’Hondt’s Algorithm for On-line Recommendations Aggregation
Fuzzy D’Hondt’s Algorithm for On-line Recommendations Aggregation
 
Off-line vs. On-line Evaluation of Recommender Systems in Small E-commerce
Off-line vs. On-line Evaluation of Recommender Systems in Small E-commerceOff-line vs. On-line Evaluation of Recommender Systems in Small E-commerce
Off-line vs. On-line Evaluation of Recommender Systems in Small E-commerce
 
Linking Content Information with Bayesian Personalized Ranking via Multiple C...
Linking Content Information with Bayesian Personalized Ranking via Multiple C...Linking Content Information with Bayesian Personalized Ranking via Multiple C...
Linking Content Information with Bayesian Personalized Ranking via Multiple C...
 
Towards Complex User Feedback and Presentation Context in Recommender Systems
Towards Complex User Feedback and Presentation Context in Recommender SystemsTowards Complex User Feedback and Presentation Context in Recommender Systems
Towards Complex User Feedback and Presentation Context in Recommender Systems
 
Using the Context of User Feedback in Recommender Systems
Using the Context of User Feedback in Recommender SystemsUsing the Context of User Feedback in Recommender Systems
Using the Context of User Feedback in Recommender Systems
 
How to Interpret Implicit User Feedback
How to Interpret Implicit User FeedbackHow to Interpret Implicit User Feedback
How to Interpret Implicit User Feedback
 
Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...
 
RecSys Challenge 2014, SemWexMFF group
RecSys Challenge 2014, SemWexMFF groupRecSys Challenge 2014, SemWexMFF group
RecSys Challenge 2014, SemWexMFF group
 

Kürzlich hochgeladen

Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...
Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...
Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...Chiheb Ben Hammouda
 
Oxo-Acids of Halogens and their Salts.pptx
Oxo-Acids of Halogens and their Salts.pptxOxo-Acids of Halogens and their Salts.pptx
Oxo-Acids of Halogens and their Salts.pptxfarhanvvdk
 
Food_safety_Management_pptx.pptx in microbiology
Food_safety_Management_pptx.pptx in microbiologyFood_safety_Management_pptx.pptx in microbiology
Food_safety_Management_pptx.pptx in microbiologyHemantThakare8
 
lect1 introduction.pptx microbiology ppt
lect1 introduction.pptx microbiology pptlect1 introduction.pptx microbiology ppt
lect1 introduction.pptx microbiology pptzbyb6vmmsd
 
Speed Breeding in Vegetable Crops- innovative approach for present era of cro...
Speed Breeding in Vegetable Crops- innovative approach for present era of cro...Speed Breeding in Vegetable Crops- innovative approach for present era of cro...
Speed Breeding in Vegetable Crops- innovative approach for present era of cro...jana861314
 
Environmental Acoustics- Speech interference level, acoustics calibrator.pptx
Environmental Acoustics- Speech interference level, acoustics calibrator.pptxEnvironmental Acoustics- Speech interference level, acoustics calibrator.pptx
Environmental Acoustics- Speech interference level, acoustics calibrator.pptxpriyankatabhane
 
Understanding Nutrition, 16th Edition pdf
Understanding Nutrition, 16th Edition pdfUnderstanding Nutrition, 16th Edition pdf
Understanding Nutrition, 16th Edition pdfHabibouKarbo
 
LAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary Microbiology
LAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary MicrobiologyLAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary Microbiology
LAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary MicrobiologyChayanika Das
 
Loudspeaker- direct radiating type and horn type.pptx
Loudspeaker- direct radiating type and horn type.pptxLoudspeaker- direct radiating type and horn type.pptx
Loudspeaker- direct radiating type and horn type.pptxpriyankatabhane
 
Abnormal LFTs rate of deco and NAFLD.pptx
Abnormal LFTs rate of deco and NAFLD.pptxAbnormal LFTs rate of deco and NAFLD.pptx
Abnormal LFTs rate of deco and NAFLD.pptxzeus70441
 
BACTERIAL SECRETION SYSTEM by Dr. Chayanika Das
BACTERIAL SECRETION SYSTEM by Dr. Chayanika DasBACTERIAL SECRETION SYSTEM by Dr. Chayanika Das
BACTERIAL SECRETION SYSTEM by Dr. Chayanika DasChayanika Das
 
DNA isolation molecular biology practical.pptx
DNA isolation molecular biology practical.pptxDNA isolation molecular biology practical.pptx
DNA isolation molecular biology practical.pptxGiDMOh
 
6.2 Pests of Sesame_Identification_Binomics_Dr.UPR
6.2 Pests of Sesame_Identification_Binomics_Dr.UPR6.2 Pests of Sesame_Identification_Binomics_Dr.UPR
6.2 Pests of Sesame_Identification_Binomics_Dr.UPRPirithiRaju
 
ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...
ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...
ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...Chayanika Das
 
AICTE activity on Water Conservation spreading awareness
AICTE activity on Water Conservation spreading awarenessAICTE activity on Water Conservation spreading awareness
AICTE activity on Water Conservation spreading awareness1hk20is002
 
Introduction of Human Body & Structure of cell.pptx
Introduction of Human Body & Structure of cell.pptxIntroduction of Human Body & Structure of cell.pptx
Introduction of Human Body & Structure of cell.pptxMedical College
 
CHROMATOGRAPHY PALLAVI RAWAT.pptx
CHROMATOGRAPHY  PALLAVI RAWAT.pptxCHROMATOGRAPHY  PALLAVI RAWAT.pptx
CHROMATOGRAPHY PALLAVI RAWAT.pptxpallavirawat456
 
3.-Acknowledgment-Dedication-Abstract.docx
3.-Acknowledgment-Dedication-Abstract.docx3.-Acknowledgment-Dedication-Abstract.docx
3.-Acknowledgment-Dedication-Abstract.docxUlahVanessaBasa
 
Combining Asynchronous Task Parallelism and Intel SGX for Secure Deep Learning
Combining Asynchronous Task Parallelism and Intel SGX for Secure Deep LearningCombining Asynchronous Task Parallelism and Intel SGX for Secure Deep Learning
Combining Asynchronous Task Parallelism and Intel SGX for Secure Deep Learningvschiavoni
 

Kürzlich hochgeladen (20)

Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...
Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...
Efficient Fourier Pricing of Multi-Asset Options: Quasi-Monte Carlo & Domain ...
 
Oxo-Acids of Halogens and their Salts.pptx
Oxo-Acids of Halogens and their Salts.pptxOxo-Acids of Halogens and their Salts.pptx
Oxo-Acids of Halogens and their Salts.pptx
 
Food_safety_Management_pptx.pptx in microbiology
Food_safety_Management_pptx.pptx in microbiologyFood_safety_Management_pptx.pptx in microbiology
Food_safety_Management_pptx.pptx in microbiology
 
lect1 introduction.pptx microbiology ppt
lect1 introduction.pptx microbiology pptlect1 introduction.pptx microbiology ppt
lect1 introduction.pptx microbiology ppt
 
Speed Breeding in Vegetable Crops- innovative approach for present era of cro...
Speed Breeding in Vegetable Crops- innovative approach for present era of cro...Speed Breeding in Vegetable Crops- innovative approach for present era of cro...
Speed Breeding in Vegetable Crops- innovative approach for present era of cro...
 
Environmental Acoustics- Speech interference level, acoustics calibrator.pptx
Environmental Acoustics- Speech interference level, acoustics calibrator.pptxEnvironmental Acoustics- Speech interference level, acoustics calibrator.pptx
Environmental Acoustics- Speech interference level, acoustics calibrator.pptx
 
Understanding Nutrition, 16th Edition pdf
Understanding Nutrition, 16th Edition pdfUnderstanding Nutrition, 16th Edition pdf
Understanding Nutrition, 16th Edition pdf
 
LAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary Microbiology
LAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary MicrobiologyLAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary Microbiology
LAMP PCR.pptx by Dr. Chayanika Das, Ph.D, Veterinary Microbiology
 
Loudspeaker- direct radiating type and horn type.pptx
Loudspeaker- direct radiating type and horn type.pptxLoudspeaker- direct radiating type and horn type.pptx
Loudspeaker- direct radiating type and horn type.pptx
 
Abnormal LFTs rate of deco and NAFLD.pptx
Abnormal LFTs rate of deco and NAFLD.pptxAbnormal LFTs rate of deco and NAFLD.pptx
Abnormal LFTs rate of deco and NAFLD.pptx
 
BACTERIAL SECRETION SYSTEM by Dr. Chayanika Das
BACTERIAL SECRETION SYSTEM by Dr. Chayanika DasBACTERIAL SECRETION SYSTEM by Dr. Chayanika Das
BACTERIAL SECRETION SYSTEM by Dr. Chayanika Das
 
DNA isolation molecular biology practical.pptx
DNA isolation molecular biology practical.pptxDNA isolation molecular biology practical.pptx
DNA isolation molecular biology practical.pptx
 
Introduction Classification Of Alkaloids
Introduction Classification Of AlkaloidsIntroduction Classification Of Alkaloids
Introduction Classification Of Alkaloids
 
6.2 Pests of Sesame_Identification_Binomics_Dr.UPR
6.2 Pests of Sesame_Identification_Binomics_Dr.UPR6.2 Pests of Sesame_Identification_Binomics_Dr.UPR
6.2 Pests of Sesame_Identification_Binomics_Dr.UPR
 
ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...
ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...
ESSENTIAL FEATURES REQUIRED FOR ESTABLISHING FOUR TYPES OF BIOSAFETY LABORATO...
 
AICTE activity on Water Conservation spreading awareness
AICTE activity on Water Conservation spreading awarenessAICTE activity on Water Conservation spreading awareness
AICTE activity on Water Conservation spreading awareness
 
Introduction of Human Body & Structure of cell.pptx
Introduction of Human Body & Structure of cell.pptxIntroduction of Human Body & Structure of cell.pptx
Introduction of Human Body & Structure of cell.pptx
 
CHROMATOGRAPHY PALLAVI RAWAT.pptx
CHROMATOGRAPHY  PALLAVI RAWAT.pptxCHROMATOGRAPHY  PALLAVI RAWAT.pptx
CHROMATOGRAPHY PALLAVI RAWAT.pptx
 
3.-Acknowledgment-Dedication-Abstract.docx
3.-Acknowledgment-Dedication-Abstract.docx3.-Acknowledgment-Dedication-Abstract.docx
3.-Acknowledgment-Dedication-Abstract.docx
 
Combining Asynchronous Task Parallelism and Intel SGX for Secure Deep Learning
Combining Asynchronous Task Parallelism and Intel SGX for Secure Deep LearningCombining Asynchronous Task Parallelism and Intel SGX for Secure Deep Learning
Combining Asynchronous Task Parallelism and Intel SGX for Secure Deep Learning
 

Towards Recommender Systems for Police Photo Lineup

  • 1. Towards Recommender Systems for Police Photo Lineup Ladislav Peškaa and Hana Trojanováb a Department of Software Engineering, b Department of Psychology, Charles University, Prague, Czech Republic
  • 2. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 2 Recommender Systems  Propose relevant items to the right persons at the right time  Successful on many various domains  Multimedia  E-commerce  POIs  Events  Recepies  …  However, some novel (relevant, important) tasks are yet to discover DLRS@RecSys2017, Como
  • 3. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 3 Police Photo Lineup  Eyewitness identification of the suspect / offender during criminal proceedings  Most common case: select suspect among other persons (fillers)  Either „in natura“ or based on photographies DLRS@RecSys2017, Como - Suspect is positioned at random - 3-7 fillers are added, witness should not know them - Wittness is instructed that the offender may or may not be present
  • 4. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 4 Errors in Photo Lineup DLRS@RecSys2017, Como 1) Observe the event 2) Identify the offender - among other candidates
  • 5. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 5 Errors in Photo Lineup DLRS@RecSys2017, Como 1) Observe the event. - numerous sources of error (poor lighting, large distance, short time, fear, fatigue…) - affects memory and recognition processes - decrease witness certainty and reliability 2) Identify the offender - decreased certainty may lead to incorrect identification - and conviction of innocent persons
  • 6. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 6 Assembling Fair Photo Lineups  System should be set to eliminate testimony of uncertain witnesses  Double-blind administration  Fair (unbiased) assembling of lineups  Out of the dataset of candidates, lineup administrator should select fillers similar to the suspect  Currently, only feature-based filtering systems are available  Better automatization in content processing is necessary  How to better approximate human-perceived similarity of persons?  Convolutional Neural Networks? DLRS@RecSys2017, Como
  • 7. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 7 Assembling Fair Photo Lineups  First approach: item-based recommendations  Propose top-k candidates for each suspect  CB-RS, cosine similarity of candidates’ and suspect’s CB features  Nationality, Age, Appearence features  Baseline (mimic feature-based filters)  Visual-RS, cosine similarity of candidates’ and suspect’s VGG- Face probability layers  Approximate the similarity of a candidate and the suspect through their similarities with other persons  VGG-Face1 network  VGG network for facial recognition problem  Trained as a multi-classification problem  Dataset of 2,622 celebrities with 1,000 images each 1Parkhi, O. M., Vedaldi, A., & Zisserman, A. (2015). Deep Face Recognition. British Machine Vision Conference. DLRS@RecSys2017, Como
  • 8. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 8 Research Questions  Is candidates recommendation (based on VGG-Face’s features) suitable for the lineup assembling?  How well is the human-perceived similarity approximated within top-k candidates?  What is the role of diversity in lineup assembling?  Is there a space for personalized recommendations (i.e. personalized per lineup administrator)?  Can there be „too similar“ candidates? DLRS@RecSys2017, Como
  • 9. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 9 Evaluation – Protocol  Dataset of missing and wanted persons in Czech Republic  4423 males  User study, 30 different lineups, 7 forensic technicians  For each lineup, both CB-RS and Visual-RS proposed top-20 candidates  Merged into one list, forensic technician should select relevant fillers for particular suspect  Follow-up questionaire and discussion DLRS@RecSys2017, Como
  • 10. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 10 Results  In total 202 completed lineups, 800 selected candidates  Recommendations were seemingly relevant in most cases  Visual-RS considerably outperformed CB-RS  The performance difference was much smaller for suspects from outside of central Europe  Effect of different ethnicity? (Known in the forensic psychology literature). DLRS@RecSys2017, Como Total lineups Selected candidates Level of agreement All lineups Visual-RS 202 466 (58%) 0.178 CB-RS 298 (37%) 0.138 Lineups with suspects from countries outside Central Europe Visual-RS 53 105 (51%) 0.128 CB-RS 82 (40%) 0.205
  • 11. Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 11 Results II.  Very low intersection (1.8%) of top-20 CB-RS and Visual-RS candidates  Potential to apply some combination in the future work  Average position of selected candidate CB-RS: 8.9 (+/- 5.2); Visual-RS: 8.2 (+/- 5.7)  Ordering seems relevant even in top-20, however can be improved  Low level of agreement among participants on the selected candidates (0.178, 0.138)  Limits of the global ordering?  Potential for personalized recommendation?  Less diversity is better  Participants agreed on the need for providing homogeneous lineups  Dynamical recommendation based on the selected candidates?  No seemingly too similar candidates  However, the dataset contained some duplicates  Should be addressed in the future work DLRS@RecSys2017, Como
  • 12. Conclusions  Recommending candidates based on CNN’s features seems plausible in police photo lineup  Item-based recommendations seems like a good first approach  There is a potential for:  Session-based personalized recommendations (need for uniformity of the final list)  Long-term preference learning (due to the level of disaggreeement among lineup administrators)  Approaches combining CNN’s features with content-based attributes Future work  Move from recommended candidates to recommended lineups  Evaluation through mock witnesses  Deployable demo DLRS@RecSys2017, Como Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 12
  • 13. DLRS@RecSys2017, Como Ladislav Peska & Hana Trojanova: Towards Recommender Systems for Police Photo Lineup 13 Thank you! Questions, comments? Supplementary materials: http://github.com/lpeska/lineups, https://tinyurl.com/yat3rtr2 Slides: https://www.slideshare.net/LadislavPeska/towards-recommender-systems-for-police-photo-lineup