Vision-Language Models Empowered Nighttime Object Detection With Consistency Sampler and Hallucination Feature...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detectors struggle with cross-domain generalization, especially in adverse conditions like nighttime.
• Existing methods face challenges with instance-level features, cross-domain representation, and category generation.

Purpose of the Study:

• To propose a domain-adaptive detection framework for robust cross-domain generalization.
• To enhance object detection performance in challenging visual environments without increasing inference overhead.

Main Methods:

• A centerness-category consistency sampler and loss for improved instance-level feature selection and alignment.
• Vision-language model (VLM)-based orthogonality enhancement for better cross-domain feature distinguishability.
• A hallucination feature generator to synthesize robust features for missing categories, ensuring balanced participation.

Main Results:

• The proposed framework consistently outperforms state-of-the-art detectors across various domain adaptation and generalization settings.
• Achieved up to 5.5 mAP improvement, demonstrating significant gains.
• Showcased particularly strong performance improvements in nighttime adaptation scenarios.

Conclusions:

• The domain-adaptive framework effectively addresses key challenges in cross-domain object detection.
• The method offers a robust solution for generalizing object detectors to new domains, including challenging nighttime conditions.
• This work contributes to more reliable and versatile object detection systems.