FeaNorm

FeaNorm is an intuitive study to compare different feature normalization strategies for part-based image classification. The part-based models used in this paper are Spatial Pyramid Matching (SPM) [Lazebnik, CVPR06] and Hierarchical Part Matching (HPM) [Xie, ICCV13]. In the context of SVM classifier, we first formulate power and coeffient, two key parameters in feature normalization, and then propose two part-based properties, i.e., the independent assumption and the hierarchical-contribution assumption. In this way, it is possible to generate more discriminative image representation with merely no extra computational overheads. The contribution of this paper is twofold:

• We verify that L1-norm does NOT result in much lower classification accuracy than L2-norm, which is a popular opinion [Boureau, CVPR10] [Wang, CVPR10].
• We propose to use part-based priors for a more intuitive and balanced normalization strategy, which does produce better classification results.

• What does the name "FeaNorm" stand for?
FeaNorm is short for Feature Normalization.


• Does FeaNorm work well in all the classification tasks?
No. It highly depends on whether the part-based assumptions (independent and hierarchical-contribution) really fit the classification model. Generally speaking, the part-based properties in SPM is much weaker than HPM, since SPM adopts a naive spatial segmentation which is less likely to generate "parts". However, it is verified that FeaNorm would not lower the classification accuracy, even if it does not produce significant improvements.


• If I want to use the codes for my experiments, what paper shall I cite?
Please cite our ICIP 2013 paper.

• For detailed guidance for running the codes, please refer to the README file in each code package.
• NEW! [RAR package][ZIP package] Version 1.0, released in September, 2013.
• NEW! [RAR package][ZIP package] "CUB-200-2011_Features", the feature-set containing all the 200 categories from Caltech-UCSD-Birds-200-2011 dataset. We have provided the RAW version of features (without normalization). You can check the usefulness of the feature normalization by comparing classification results before and after normalization.
• If you find any mistakes or bugs in the codes or datasets, please contact me.

• Lingxi Xie, Qi Tian and Bo Zhang, "Feature Normalization for Part-Based Image Classification", in IEEE International Conference on Image Processing (ICIP), Oral Presentation (Acceptance Rate ~ 15%), pages 2607--2611, Melbourne, Australia, 2013. [PDF] [Slides] [BibTeX]

• [Xie, ICCV13] Lingxi Xie, Qi Tian, Shuicheng Yan and Bo Zhang, "Hierarchical Part Matching for Fine-Grained Visual Categorization", in IEEE International Conference on Computer Vision (ICCV), 2013.
• [Wang, CVPR10] Jinjun Wang, Jianchao Yang, Kai Yu, Fengjun Lv, Thomas Huang, and Yihong Gong, "Locality-Constrained Linear Coding for Image Classification", in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010.
• [Boureau, CVPR10] Y-Lan Boureau, Francis Bach, Yann LeCun and Jean Ponce, "Learning Mid-Level Features for Recognition", in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010.
• [Lazebnik, CVPR06] Svetlana Lazebnik, Cordelia Schmid and Jean Ponce, "Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories", in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2006.