Fast and accurate RFD-like descriptor approximation for SIMD architectures

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Local image descriptors are crucial for computer vision tasks but can be computationally intensive.
• Comparing descriptors is time-consuming for large datasets, posing challenges for real-time and mobile applications.
• Binary descriptors offer efficient comparison via Hamming distance, but existing accurate ones like RFD and RFDoc are computationally demanding on edge devices.

Purpose of the Study:

• To develop a computationally efficient approximation of RFD-like image descriptors.
• To accelerate descriptor computation for real-time and mobile computer vision applications.
• To evaluate the accuracy and efficiency of the proposed descriptor.

Main Methods:

• Extended the Fast and Accurate RFD-like descriptor Approximation (FARA) method.
• Adapted FARA to utilize Single Instruction Multiple Data (SIMD) CPU extensions for all descriptor computation stages.
• Implemented CPU-friendly, integer-only gradient map computation and sequential memory accesses.

Main Results:

• Achieved a twofold acceleration in gradient map computation.
• Reduced overall descriptor computation time by half compared to the fast RFDoc implementation on x86 and ARM architectures.
• Demonstrated comparable accuracy to RFDoc on MIDV datasets and RFD on Brown and Oxford datasets for document localization and classification.

Conclusions:

• FARA offers a significant speedup in image descriptor computation without compromising accuracy.
• The SIMD-adapted FARA is suitable for resource-constrained environments and real-time computer vision applications.
• FARA presents a viable alternative to existing descriptors for tasks requiring efficient and accurate image feature identification.