Download presentation

Presentation is loading. Please wait.

Published byRussell Savery Modified over 2 years ago

0
**Relevance Feedback for the Earth Mover‘s Distance**

Marc Wichterich, Christian Beecks, Martin Sundermeyer, Thomas Seidl Data Management and Data Exploration Group RWTH Aachen University, Germany

1
**Introduction Distance-based Adaptable Similarity Search**

Similarity of objects defined by distance function Small distance → similar, large distance → dissimilar Query by example: user-given object, find similar ones Query and distance only approximate descriptions of user’s desired result If delivered result does not meet expectations: Bad query? Bad distance? Bad database? How to do it better? Relevance Feedback attempts to adapt query/similarity model based on simple user input (result relevancy)

2
**RF for EMD Relevance Feedback Earth Mover’s Distance RF EMD**

user DB query, feedback results feedback system similarity model Photo: Flickr / Caro Wallis Earth Mover’s Distance RF for EMD RF EMD

3
**Overview Introduction Adaptive Similarity Model**

Feature Signatures The Earth Mover’s Distance Relevance Feedback for the Earth Mover’s Distance Experimental Evaluation Conclusion

4
**Similarity Model – Feature Signatures**

color x

5
**Similarity Model – Earth Mover’s Distance**

Introduced in Computer Vision by Rubner et al. Used in many differing application domains Idea: transform features of Q into features of P EMD: minimum of transformation cost Q P x y x y

6
**Feature Transformation**

7
**EMD – Formal Definition**

Modeled as linear optimization (transportation problem)

8
**Overview Introduction Adaptive Similarity Model**

Relevance Feedback for the Earth Mover’s Distance The Feedback Loop Query Adaptation Heuristic EMD Adaptation Optimization-based EMD Adaptation Experimental Evaluation Conclusion

9
**? The Feedback Loop no yes exit start get query retrieve results**

user DB query, feedback results feedback system similarity model start get query retrieve results adapt distance ? display results adapt query satisfied? no get feedback yes exit

10
**Query Adaptation Input: signatures from relevant objects**

satisfied? retrieve results exit start get query query distance feedback display results Query Adaptation Input: signatures from relevant objects Output: new query signature Idea: cluster signature elements Refinements by Rubner: Only keep clusters with elements from majority of signatures Reweight resulting signature accordingly Combine with fixed gd L2 and call it „Query-by-Refinement“ „Query-by-Refinement“ is baseline for our evaluation We adapt EMD via ground distance

11
**Heuristic EMD Adaptation 1**

satisfied? retrieve results exit start get query query distance feedback display results Heuristic EMD Adaptation 1 Approach: pick gd based on feedback gd should reflect user preferences: Don’t care if blue cluster at upper half of image is moved left/right Do care if it is moved vertically Use variance information in relevant feedback Low variance → assume user cares High variance → assume user does not care Measure variance in feedback locally around query signature elements ci(Q). Define gd: c(Q) x FS → R ( )

12
**Heuristic EMD Adaptation 2**

satisfied? retrieve results exit start get query query distance feedback display results Heuristic EMD Adaptation 2 Not 1 but m distance functions: gdi(ci(Q),y) = ((ci(Q)- y) Vi (ci(Q)- y)T)½ Weighted Euclidean Distances (weights on diagonal of Vi) Vi : inverted variance for ci(Q) per feature space dimension

13
**Optimization-Based EMD Adaptation 1**

satisfied? retrieve results exit start get query query distance feedback display results Optimization-Based EMD Adaptation 1 Aim: Pick best possible gd. Failback: Find a good one. Q: When is gd good? A: If ranking it produces is good. New Q: When is a ranking of DB good? Given ground truth, a number of measures exist We used “average precision at relevant positions” We have ground truth for part of the DB: feedback Idea: test candidates for gd on feedback Ranking Avg. Precision 1.000 0.854 0.365

14
**Optimization-Based EMD Adaptation 2**

satisfied? retrieve results exit start get query query distance feedback display results Optimization-Based EMD Adaptation 2 Optimization: Optimization variable: gd Objective function: avgPrec(EMDgd , q, Feedback) Constraints: m weighted Euclidean distances Analytic optimization with closed form for weights infeasible (ranking/sorting, EMDs in objective function) Probabilistic optimization via Simulated Annealing Start with some initial solution Move in solution space Compute objective function Adopt solution with certain probability Iterate & turn more greedy

15
**Optimization-Based EMD Adaptation 3**

satisfied? retrieve results exit start get query query distance feedback display results Optimization-Based EMD Adaptation 3 Optimization for EMD based on Feedback: Solution: weights for m weighted Euclidean distances Initial solution: given by heuristic Moving: redistribute weights per Euclidean distance Objective function: avgPrec(EMDgd , q, Feedback) Results for EMDgd on DB?

16
**Overview Introduction Adaptive Similarity Model**

Relevance Feedback for the Earth Mover’s Distance Experimental Evaluation Conclusion

17
**Experimental Evaluation: Databases**

72,000 images in ALOI DB ~60,000 images in COREL DB

18
**Experimental Evaluation: ALOI**

Query-by-Refinement Heuristic Adaptation Optimization-based

19
**Experimental Evaluation: COREL**

Query-by-Refinement Heuristic Adaptation Optimization-based

20
**Experimental Evaluation**

After 5 iterations of looking for doors in COREL: (a) Query-by-Refinement (b) Heuristic (c) Optimization-Based pos

21
**Conclusion Exploited adaptability of the EMD in RF framework**

Goal: Improve similarity search results Techniques: Baseline: fixed ground distance Statistics-based heuristic adaptation Optimization-based adaptation Evaluation: Experiments on two image datasets More relevant objects in fewer iterations Techniques extensible to other adaptable distance functions

Similar presentations

OK

Computer Vision Group, University of BonnVision Laboratory, Stanford University Abstract This paper empirically compares nine image dissimilarity measures.

Computer Vision Group, University of BonnVision Laboratory, Stanford University Abstract This paper empirically compares nine image dissimilarity measures.

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google

Maths ppt on number system for class 9 Ppt on bluetooth based smart sensor Ppt on viruses and anti viruses for free Ppt on sports day outfits Ppt on hindu religion map Ppt on power factor correction using capacitor bank Ppt on idiopathic thrombocytopenia purpura in children Hrm ppt on recruitment 2016 Ppt on pricing policy of a company Free download ppt on conservation of wildlife