Compositional Model Based Fisher Vector Coding for Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Fisher vector coding (FVC) is effective for image classification but often limited by Gaussian mixture models (GMMs).
• GMMs assume a fixed, small number of prototypes, restricting their representative power for local features.

Purpose of the Study:

• To enhance the representative power of generative models for FVC.
• To introduce novel FVC strategies that overcome GMM limitations.

Main Methods:

• Developed two compositional generative models for local features, moving beyond traditional GMM assumptions.
• Proposed Sparse Coding-based FVC (SCFVC) and Hybrid SCFVC (HSCFVC) derived from these models.
• Utilized pre-trained Convolutional Neural Network (CNN) activations as local features.

Main Results:

• The proposed SCFVC and HSCFVC methods significantly outperform traditional GMM-based FVC.
• Achieved state-of-the-art performance on various image classification tasks.
• Demonstrated the superiority of compositional generative models in FVC.

Conclusions:

• Novel FVC strategies (SCFVC, HSCFVC) offer enhanced representation capabilities.
• These methods provide a powerful alternative to GMM-based FVC for image classification.
• The compositional approach advances the field of feature coding in machine learning.