SRCD: Semantic Reasoning With Compound Domains for Single-Domain Generalized Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Single-domain generalized object detection (Single-DGOD) is challenging due to limited source data.
• Existing methods struggle with pseudo attribute-label correlations and ignore crucial semantic structure information.
• Instance-level semantic relations are vital for robust model generalization.

Purpose of the Study:

• To propose a novel framework, Semantic Reasoning with Compound Domains (SRCD), for Single-DGOD.
• To enhance model generalization by learning and maintaining semantic structures of self-augmented compound cross-domain samples.
• To address limitations of existing methods in handling scarce single-domain data and semantic structure information.

Main Methods:

• Introduced Semantic Reasoning with Compound Domains (SRCD) framework for Single-DGOD.
• Developed a texture-based self-augmentation (TBSA) module to eliminate irrelevant attribute effects.
• Implemented a local-global semantic reasoning (LGSR) module to model instance-level semantic relationships and preserve intrinsic structures.

Main Results:

• The proposed SRCD framework significantly improves generalization ability in Single-DGOD tasks.
• Experiments on multiple benchmarks validate the effectiveness of SRCD in enhancing object detection performance.
• The TBSA and LGSR modules effectively address limitations of previous approaches.

Conclusions:

• The SRCD framework offers a promising solution for the challenging Single-DGOD problem.
• Learning and maintaining semantic structures through self-augmentation and reasoning is key to robust generalization.
• The proposed approach advances the state-of-the-art in domain generalized object detection.