Download presentation
Presentation is loading. Please wait.
Published byBryan Greer Modified over 6 years ago
1
IEEE ICIP 2013 Feature Normalization for Part-Based Image Classification
Speaker: Lingxi Xie Authors: Lingxi Xie, Qi Tian, Bo Zhang State Key Laboratory of Intelligent Technology and Systems Department of Computer Science and Technology Tsinghua University
2
ICIP 2013 - Oral Presentation
Outline Image Classification The Part-Based Classification Models Feature Normalization Global Normalization Separate Normalization Hierarchical Normalization Experimental Results Conclusions 11/18/2018 ICIP Oral Presentation
3
ICIP 2013 - Oral Presentation
Outline Image Classification The Part-Based Classification Models Feature Normalization Global Normalization Separate Normalization Hierarchical Normalization Experimental Results Conclusions 11/18/2018 ICIP Oral Presentation
4
ICIP 2013 - Oral Presentation
Image Classification A basic task towards image understanding General vs. Fine-Grained 11/18/2018 ICIP Oral Presentation
5
ICIP 2013 - Oral Presentation
Outline Image Classification The Part-Based Classification Models Feature Normalization Global Normalization Separate Normalization Hierarchical Normalization Experimental Results Conclusions 11/18/2018 ICIP Oral Presentation
6
Spatial Pyramid Matching (SPM)
= Part 1 [Lazebnik, CVPR06] = Part 2 = Part 3 = Part 4 = Part 5 11/18/2018 ICIP Oral Presentation
7
Hierarchical Part Matching (HPM)
= Part 1 [Xie, ICCV13] = Part 2 = Part 3 = Part 4 = Part 5 11/18/2018 ICIP Oral Presentation
8
ICIP 2013 - Oral Presentation
Outline Image Classification The Part-Based Classification Models Feature Normalization Global Normalization Separate Normalization Hierarchical Normalization Experimental Results Conclusions 11/18/2018 ICIP Oral Presentation
9
Feature Normalization
A Key Step before Classifiers Data pre-processing. Equivalent the range and weight of input vectors. Significantly impact the classification results. Different Models, Different Normalization. Support Vector Machine (SVM) Naïve Bayes Classificer (NB) Hidden Markov Models (HMM) 11/18/2018 ICIP Oral Presentation
10
ICIP 2013 - Oral Presentation
Global Normalization Considering the Feature as a Whole. coefficient norm 11/18/2018 ICIP Oral Presentation
11
ICIP 2013 - Oral Presentation
Global Normalization head head body body black-footed albatross sooty albatross 11/18/2018 ICIP Oral Presentation
12
ICIP 2013 - Oral Presentation
Global Normalization 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL not balanced 11/18/2018 ICIP Oral Presentation
13
FAIR? Global Normalization Small Parts have Low Feature Weights. 1 2 3
4 5 6 head body ALL 1 2 3 4 5 6 head body ALL FAIR? 11/18/2018 ICIP Oral Presentation
14
beak wing Global Normalization
However, Small Parts are also Important. groove billed ani common raven red winged blackbird rusty blackbird beak wing 11/18/2018 ICIP Oral Presentation
15
Separate Normalization
Normalizing each Part Individually part-wise vector coefficient 11/18/2018 ICIP Oral Presentation
16
Separate Normalization
1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL Small Parts are Given Equal Weights! 11/18/2018 ICIP Oral Presentation
17
Separate Normalization
Different-Level Parts have Same Weight. 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL FAIR? 11/18/2018 ICIP Oral Presentation
18
Separate Normalization
Examples of Hierarchical Parts in Birds Dataset. head = beak + eyes + crown + forehead neck = nape + throat body = breast + back + belly + wings ALL = head + body + tail + legs Some Observations: High-level parts contain more information. High-level parts are less likely to be missing. 11/18/2018 ICIP Oral Presentation
19
Hierarchical Normalization
Assigning more Weights on High-Level Parts hierarchical contribution part-wise weight part-wise coefficient 11/18/2018 ICIP Oral Presentation
20
Hierarchical Normalization
1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL 1 2 3 4 5 6 head body ALL High-Level Parts are Enhanced! 11/18/2018 ICIP Oral Presentation
21
ICIP 2013 - Oral Presentation
Outline Image Classification The Part-Based Classification Models Feature Normalization Global Normalization Separate Normalization Hierarchical Normalization Experimental Results Conclusions 11/18/2018 ICIP Oral Presentation
22
ICIP 2013 - Oral Presentation
The Caltech101 Dataset General Object Recognition [Fei-Fei, CVIU07] 102 classes (one background category) 9144 images Models SPM [Lazebnik, CVPR06] + LLC [Wang, CVPR10] SPM [Lazebnik, CVPR06] + GPP [Xie, MM12] SPM [Lazebnik, CVPR06] + EdgeGPP [Xie, MM12] 16 Basic Parts + 5 High-Level Parts 11/18/2018 ICIP Oral Presentation
23
ICIP 2013 - Oral Presentation
The Caltech101 Dataset SPM+LLC SPM+GPP SPM+EdgeGPP No Normalization 73.14 76.35 80.78 Global L1-norm 73.91 76.26 80.86 Global L2-norm 74.41 77.03 82.45 Global Li-norm 73.25 76.47 80.89 Separate L1-norm 71.99 75.20 78.05 Separate L2-norm 73.68 75.47 81.24 Separate Li-norm 73.39 76.43 80.93 Hierarchical L1-norm 72.71 75.88 80.40 Hierarchical L2-norm 74.31 76.86 83.19 Hierarchical Li-norm 73.89 76.55 81.37 11/18/2018 ICIP Oral Presentation
24
ICIP 2013 - Oral Presentation
The CUB Dataset Fine-Grained Bird Classification [Wah, TR11] 200 species of birds 11788 images Models Parts [Xie, ICCV13] + LLC [Wang, CVPR10] HPM [Xie, ICCV13] + LLC [Wang, CVPR10] HPM [Xie, ICCV13] + GPP [Xie, MM12] 15 Basic Parts + 6 High-Level Parts 11/18/2018 ICIP Oral Presentation
25
ICIP 2013 - Oral Presentation
The CUB Dataset Parts+LLC HPM+LLC HPM+GPP No Normalization 25.58 27.55 30.67 Global L1-norm 27.61 29.85 33.22 Global L2-norm 27.06 29.30 32.96 Global Li-norm 25.04 27.41 30.71 Separate L1-norm 24.58 26.94 31.84 Separate L2-norm 30.93 32.75 35.98 Separate Li-norm 27.73 29.67 31.92 Hierarchical L1-norm - 28.12 33.85 Hierarchical L2-norm - 32.89 36.48 Hierarchical Li-norm - 29.45 32.08 11/18/2018 ICIP Oral Presentation
26
ICIP 2013 - Oral Presentation
Discussions The Performance of Our Model SPM: only comparable with the original model. HPM: significantly better! Why? Both Assumptions Sound? There exist more semantics in HPM! 11/18/2018 ICIP Oral Presentation
27
ICIP 2013 - Oral Presentation
Outline Image Classification The Part-Based Classification Models Feature Normalization Global Normalization Separate Normalization Hierarchical Normalization Experimental Results Conclusions 11/18/2018 ICIP Oral Presentation
28
ICIP 2013 - Oral Presentation
Conclusions Feature Normalization An important issue in image representation. In Part-Based Classification Models Instructive to consider each part separately. 3 Normalization Strategies Global Normalization Separate Normalization Hierarchical Normalization Easy to Implement! 11/18/2018 ICIP Oral Presentation
29
ICIP 2013 - Oral Presentation
Thank you! Questions please? 11/18/2018 ICIP Oral Presentation
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.