Receptive fields selection for binary feature description.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Local image patch descriptors are crucial in computer vision.
• Traditional methods rely on expert knowledge, while learning-based approaches offer data-driven advantages.
• Developing efficient and effective binary feature descriptors remains an active research area.

Purpose of the Study:

• To propose a novel data-driven method for designing binary feature descriptors.
• To introduce the receptive fields descriptor (RFD) and evaluate its performance.
• To compare RFD against state-of-the-art binary and float-valued descriptors.

Main Methods:

• Developed a data-driven approach to construct binary feature descriptors called receptive fields descriptor (RFD).
• RFD is created by thresholding responses from selected receptive fields, chosen greedily based on distinctiveness and correlation.
• Two variants, RFDR and RFDG, were generated using rectangular and Gaussian pooling areas, respectively.

Main Results:

• Image matching experiments showed RFD significantly outperforms current binary descriptors.
• RFD performance was comparable to top float-valued descriptors but with reduced processing time.
• Object recognition tasks confirmed RFD's ability to close the performance gap with float-valued descriptors.

Conclusions:

• The proposed receptive fields descriptor (RFD) offers a powerful and efficient alternative for binary feature description.
• RFD variants (RFDR, RFDG) demonstrate strong performance in image matching and object recognition.
• This data-driven method advances the field of binary feature descriptors in computer vision.