ftrdbj semantic indexing systems for trecvid 2010
play

FTRDBJ Semantic Indexing Systems for TRECVID 2010 Kun TAO France - PowerPoint PPT Presentation

FTRDBJ Semantic Indexing Systems for TRECVID 2010 Kun TAO France Telecom (R&D) Orange Labs, Beijing Nov. 15, 2010 research & development Confidential Overview 2009 HLFE Systems 7 CEGL features & 6 SIFT features 3


  1. FTRDBJ Semantic Indexing Systems for TRECVID 2010 Kun TAO France Telecom (R&D) Orange Labs, Beijing Nov. 15, 2010 research & development Confidential

  2. Overview  2009 HLFE Systems  7 CEGL features & 6 SIFT features  3 late fusion runs & 3 early fusion runs  2010 SIN Systems  7 CEGL features & 12 features based on local descriptor  3 late fusion runs & 1 early fusion run  30 concept “FT - 30” corpus  A cross-domain run research & development Confidential France Telecom Group

  3. Overview  4 runs ID TYPE DESCRIPTION MAP classifier-level-combination of 19 low- 1 F_A 0.070 level feature SVMs with equal weights linear weighted combination of 19 2 F_A feature SVMs through logistic 0.075 regression cross-domain fusion between the 3 F_C results of run_2 and the results of 05-09 0.070 TRECVID models kernel-level-combination of 14 low-level 4 L_A features with equal weighted multiple 0.063 kernel learning research & development Confidential France Telecom Group

  4. Overview  FT-30  Airplane_Flying*, Boat_Ship*, Bus*, Cityscape*, Classroom*, Demonstration_Or_Protest*, Hand*, Nighttime*, Singing*, Telephones*  Animal + , Dark-skinned_People + , Flowers + , Running + , Sitting_Down + ,  Anchorperson, Beach, Bicycles, Cats, Chair, Charts, Construction_Vehicles, Crowd, Female_Person, House_Of_Worship, Instrumental_Musician, Laboratory, Roadway_Junction, Shopping_Mall, Sports, . research & development Confidential France Telecom Group

  5. Features  7 CEGL  Color Auto-Correlograms (CAC), Color Coherence Vector (CCV), Grid Color Moments (GCM), Edge Coherence Vector (ECV), Edge Direction Histogram (EDH), Gabor feature (Gabor) and Local Binary Patterns (LBP)  12 local descriptor features  SIFT, Dense-SIFT, SIFT-no_orientation  Pyramid HOW, PLSA  Soft -Assignment  HOG research & development Confidential France Telecom Group

  6. Features  PLSA P(z) P(z|d) P(w|z) z w d research & development Confidential France Telecom Group

  7. Features  Soft – Assignment 1/ ( * ) ni Dist   ni Weight ni 1,2,3 ni 3  (1/ ( * )) i Dist i  i 1  HOG 124-D Pyramid Histograms Descriptors "Object Detection using Histograms of Oriented Gradients". http://www.pascal- network.org/challenges/VOC/voc2006/slides /dalal.pdf. Jianxiong Xiao et al. "SUN Database: Large-scale Scene Recognition from Abbey to Zoo",CVPR 2010 research & development Confidential France Telecom Group

  8. Features  MAP of different features  (60% of dev. dataset for training SVM, 40% for evaluation) MAP Group Name Feature Name Dim. 0.117 512 SIFT.HOW 0.138 2560 SIFT.2L-PHOW 0.118 512 SIFT. 3L-PHOW-PLSA S6 0.166 512 DENSE-SIFT.HOW 0.169 2560 DENSE-SIFT.2L-PHOW 0.178 512 DENSE-SIFT. 3L-PHOW-PLSA 0.134 512 SIFT.HOW-SOFT 0.148 512 SS3 SIFT-NO-ORIENTATION. HOW-SOFT 0.167 512 DENSE-SIFT. HOW-SOFT research & development Confidential France Telecom Group

  9. Features  MAP of different features MAP Group Name Feature Name Dim. 0.051 Color Auto-Correlograms (CAC) 256 0.083 Color Coherence Vector (CCV) 360 0.041 Grid Color Moments (GCM) 108 0.035 CEGL Edge Coherence Vector (ECV) 320 0.047 Edge Direction Histogram (EDH) 365 0.037 Gabor feature (Gabor) 240 0.051 Local Binary Patterns (LBP) 256 0.127 HOG.HOW 512 0.133 H3 HOG.2L-PHOW 2560 0.129 HOG. 3L-PHOW-PLSA 512 research & development Confidential France Telecom Group

  10. Basic Structure  2-Step Late Fusion  Kernel-level early fusion research & development Confidential France Telecom Group

  11. Unified Weights  Motivation  Hard to evaluation all 130 concepts × 19 features  Supported by internal evaluation  LIBLINEAR were used in all modules of 2-step fusion research & development Confidential France Telecom Group

  12. Unified Weights  Results  60% for SVM, 20% for LR, 20% for evaluation research & development Confidential France Telecom Group

  13. Unified Weights  Our best run  Something more about generalization problem research & development Confidential France Telecom Group

  14. Cross-domain  Data level & classifier level  60% for SVM, 20% for weights, 20% for evaluation research & development Confidential France Telecom Group

  15. Conclusion & Future works  Using unified weights is a valuable choice  The balance between feature numbers and computation cost  Need further research on cross-domain research & development Confidential France Telecom Group

  16. Thanks! Any questions? research & development Confidential France Telecom Group

Recommend


More recommend