Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Fine-grained Image Classification

Avatar for Yasser Souri Yasser Souri
September 16, 2014

Fine-grained Image Classification

Ms seminar presentation, Sharif University of Technology

Avatar for Yasser Souri

Yasser Souri

September 16, 2014
Tweet

More Decks by Yasser Souri

Other Decks in Research

Transcript

  1. Fine-grained Image Classification Yasser Souri Supervisor: Prof. Shohreh Kasaei Image

    Processing Lab Computer Engineering Department Sharif University of Technology
  2. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 2   Image Processing Lab - Sharif
  3. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 3   Image Processing Lab - Sharif
  4. Fine-grained Image Classification •  Image classification, when considered classes are

    all subclasses of a certain class •  They are similar in appearance Image Processing Lab - Sharif 8  
  5. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 9   Image Processing Lab - Sharif
  6. Image Classification Training/Testing Image Processing Lab - Sharif 10  

    •  Given –  Positive training images containing an object class –  Negative training images not containing that object class –  Classify a test image whether it contains the object class
  7. Vocabulary Based Methods •  Bag of words Image Processing Lab

    - Sharif 13   Feature Extraction Coding Pooling Using  keypoint  detec:on  methods  like  Harris-­‐Affine  or  densely   Using  descriptors  like  SIFT   Using  the  Vocabulary  and  vector  quan:za:ons  methods  like  KMeans   Average  pooling:  Average  of  the  codes  
  8. Vocabulary Based Methods •  Spatial Pyramid Matching Image Processing Lab

    - Sharif 14   Coding Sub-regions Pooling Pooling Concatenating
  9. Vocabulary Based Methods •  Overview Image Processing Lab - Sharif

    15   Feature Extraction Coding Pooling Concatenating
  10. Deep Learning Methods •  Network architecture •  Trained with 1.2

    million images Image Processing Lab - Sharif 16  
  11. Deep Learning Methods •  Feature engineering vs. feature learning • 

    Results of Pascal VOC 2007 [47, 23] Image Processing Lab - Sharif 18   Method   mAP   SIFT  +  VQ  +  SVM   46.54   SIFT  +  LLC  +  SVM   57.60   SIFT  +  FV  +  SVM   61.69   CNN  +  SVM   73.90  
  12. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 19   Image Processing Lab - Sharif
  13. Fine-grained Classification Methods •  Human In the Loop Methods • 

    CUB-2011 Accuracy [38] –  0 Q : 19% –  20 Q: 50% •  Not fully automatic Image Processing Lab - Sharif 25  
  14. Finding Discriminative Parts •  Is Hard and Computationally expensive • 

    Results in low accuracies •  CUB-2011 Accuracy – 41.01 % [41] Image Processing Lab - Sharif 28  
  15. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 30   Image Processing Lab - Sharif
  16. Baseline Methods •  Based on [47] •  Use the network

    for features •  Use Linear SVM for classification Image Processing Lab - Sharif 31  
  17. Baseline Methods •  Ours-1 – Not using the bounding boxes – 45.44

    % •  Ours-2 – Cropping with bounding boxes – 55.62 % •  Ours-3 – Ours-1 + Ours-2 – 57.83 % Image Processing Lab - Sharif 32  
  18. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 34   Image Processing Lab - Sharif
  19. Experimental Analysis •  Ours-4 – Model averaging – 58.00 % •  Ours-2

    vs. Ours-1 ( + 10.18 %) – Removing background helps Image Processing Lab - Sharif 35  
  20. Experimental Analysis 42.69   57.31   Image Processing Lab -

    Sharif 36   Total Agreement in Our-1, Our-2, Our-3
  21. Experimental Analysis 35.85   7.11   57.31   Image Processing

    Lab - Sharif 37   Total Agreement in Our-1, Our-2, Our-3 Agreement is 84 % correct
  22. Outline •  Introduction •  Review – Image Classification – Fine-grained Image Classification

    •  Baseline Methods •  Experimental Analysis •  Future Work 38   Image Processing Lab - Sharif
  23. Future Work 1 •  Getting the same baseline methods on

    other datasets Image Processing Lab - Sharif 39  
  24. Future Work 3 •  Finding the complementary model •  Using

    mixture of experts Image Processing Lab - Sharif 41