A proposal of prior probability-oriented clustering in feature encoding strategies

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Codebook-based feature encodings are standard in image recognition, often using k-means or Gaussian Mixture Models (GMM).
• Traditional methods face limitations with large codebooks, as clusters can converge, reducing unique representations and impacting recognition performance.
• The number of unique clusters can become smaller than the designated codebook size, hindering model effectiveness.

Purpose of the Study:

• To address the disadvantage of traditional clustering frameworks in image recognition when using large codebooks.
• To present a new clustering framework that optimizes prior probability distributions for improved feature encoding.
• To enhance the performance of image recognition systems by refining the codebook construction process.

Main Methods:

• Developed a novel clustering framework with two alternating objectives, applicable to both k-means and GMM.
• Evaluated the framework using synthetic clustering datasets to compare with traditional methods.
• Tested the approach on image recognition tasks, specifically using Fisher Vector (FV) with GMM and Vector of Locally Aggregated Descriptors (VLAD) with k-means on Birds and Butterflies datasets.

Main Results:

• The proposed approach, when alternated with k-means, showed similar results but allowed finer cluster tuning.
• Alternating the approach with GMM significantly improved objective functions and created more appropriate clusters, especially for large, complex datasets.
• While VLAD performance slightly decreased, the Fisher Vector (FV) encoding showed improved recognition performance, particularly with larger codebook sizes.

Conclusions:

• The proposed clustering framework offers improvements over traditional methods, especially when integrated with GMM for feature encoding.
• The enhanced GMM approach is particularly effective for handling large and complex datasets in image recognition.
• The study highlights the potential of refining codebook construction for better performance in advanced image recognition techniques like Fisher Vectors.