Randomized Spectrum Transformations for Adapting Object Detector in Unseen Domains.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain generalization (DG) aims to create models that perform well on unseen data by training on multiple source domains.
• Object detection faces challenges in domain generalization, termed domain generalizable object detection (DGOD).
• Existing DG methods often suffer from domain bias, leading to overfitting on specific source domains.

Purpose of the Study:

• To develop a novel method for domain generalizable object detection (DGOD).
• To address the issue of domain bias in existing DG methods for object detection.
• To improve the generalization capability of object detectors to unseen target domains.

Main Methods:

• Propose Meta Learning on Randomized Transformations (MLRT) for DGOD.
• Introduce a randomized spectrum transformation (RST) module to increase source domain diversity by randomizing frequency-space information.
• Incorporate a gradient weighting (GW) module to balance gradients across domains and alleviate domain bias.

Main Results:

• The proposed MLRT model, integrating RST and GW, demonstrates effectiveness in DGOD.
• Experiments on six benchmarks show that MLRT achieves state-of-the-art (SOTA) performance.
• The RST module successfully transforms domains, increasing data diversity, while GW balances domain gradients.

Conclusions:

• MLRT effectively learns domain-invariant object detectors.
• The combination of RST and GW significantly reduces domain bias, improving generalization.
• The proposed approach sets a new standard for domain generalizable object detection.