Fourier-KAN: Feature Distribution Decomposition and Recombination for Unknown-Domain Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Single-domain Generalized Object Detection (Single-DGOD) aims to adapt object detectors to new, unseen domains.
• A key challenge is generalizing from a single source domain to multiple diverse target domains.
• Existing methods struggle with the inherent data distribution differences across domains.

Purpose of the Study:

• To propose a novel feature recombination method to expand source domain data distributions.
• To enhance the generalization capability of object detectors to unknown domains.
• To improve detection accuracy and maintain real-time performance in cross-domain scenarios.

Main Methods:

• Fourier-KAN Feature Recombination utilizes Fast Fourier Transform (FFT) to decompose features into amplitude and phase.
• Applies the Kolmogorov-Arnold theorem to decompose components into base distributions.
• Generates diverse recombined features through multi-level recombination to emulate cross-domain variations.

Main Results:

• The proposed method effectively emulates deep cross-domain variations at the feature level.
• Demonstrates strong adaptability to both two-stage and single-stage object detection frameworks.
• Achieves outstanding detection accuracy and significantly enhances generalization on Diverse Weather and Real-to-Art benchmarks.

Conclusions:

• Fourier-KAN Feature Recombination strengthens model generalization ability to unknown domains.
• The approach maintains excellent real-time performance while improving detection accuracy.
• The method offers a promising solution for cross-domain generalization in object detection.