BSIFT: toward data-independent codebook for large scale image search.

Area of Science:

• Computer Vision
• Machine Learning
• Image Retrieval

Background:

• The Bag-of-Words (BoWs) model using Scale Invariant Feature Transform (SIFT) is prevalent in large-scale image retrieval.
• Traditional feature quantization methods in BoWs models face challenges like codebook training, reliability, and update inefficiency.

Purpose of the Study:

• To propose a novel feature quantization scheme for efficient and discriminative SIFT descriptor representation.
• To develop a method that avoids explicit codebook training and is collection-independent.

Main Methods:

• Introduced Binary SIFT (BSIFT), a method to quantize SIFT descriptors into bit-vectors.
• Utilized the first 32 bits of BSIFT as codewords for inverted file indexing.
• Implemented techniques including feature filtering, codeword expansion, and query-sensitive mask shielding to reduce quantization error.

Main Results:

• The proposed BSIFT approach demonstrates collection independence and suitability for inverted file structures.
• Experimental results on public datasets show improved index efficiency and retrieval accuracy for large-scale image search.
• The method performs effectively in resource-limited scenarios without requiring explicit codebook training.

Conclusions:

• BSIFT offers an efficient and reliable alternative to traditional feature quantization for image retrieval.
• The approach enhances scalability and accuracy in large-scale image search applications.
• BSIFT's collection independence and lack of codebook requirement broaden its applicability, especially in constrained environments.