SIFT Meets CNN: A Decade Survey of Instance Retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Early content-based image retrieval (CBIR) relied on global features.
• Local descriptors like SIFT dominated image retrieval for over a decade due to robustness.
• Convolutional Neural Network (CNN) based image representations have recently shown impressive performance.

Purpose of the Study:

• To provide a comprehensive survey of instance retrieval methods over the last decade.
• To categorize and analyze SIFT-based and CNN-based instance retrieval techniques.
• To offer insights into the evolution and future directions of instance retrieval.

Main Methods:

• Categorization of SIFT-based methods by codebook size (large, medium, small).
• Discussion of CNN-based methods: pre-trained, fine-tuned, and hybrid approaches.
• Analysis of patch-based feature extraction in CNN methods versus single-pass image processing.

Main Results:

• Milestones in modern instance retrieval are presented.
• A broad selection of previous works across different categories is reviewed.
• Comparative analysis of retrieval performance across datasets for SIFT and CNN methods.

Conclusions:

• CNN-based methods have gained significant traction and demonstrated high performance.
• Understanding the connection between SIFT and CNN methods is crucial.
• Promising future directions include generic and specialized instance retrieval.