Boosting Night-Time Scene Parsing With Learnable Frequency.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing daytime scene parsing methods fail at night due to poor illumination.
• Night-time scenes lack reliable spatial contextual cues for traditional methods.

Purpose of the Study:

• To develop a novel approach for accurate night-time scene parsing.
• To address the limitations of current methods in low-light conditions.

Main Methods:

• Statistical analysis of image frequency distributions to identify day-night discrepancies.
• Development of a Learnable Frequency Encoder (LFE) to dynamically model frequency components.
• Introduction of a Spatial Frequency Fusion (SFF) module to integrate spatial and frequency information.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches on benchmark datasets (NightCity, NightCity+, BDD100K-night).
• The approach effectively leverages frequency information for improved night-time scene understanding.
• The method enhances the performance of existing daytime scene parsing models when applied to night-time data.

Conclusions:

• Image frequency distribution is a critical factor for effective night-time scene parsing.
• The proposed LFE and SFF modules offer a robust solution for low-light vision challenges.
• This work provides a valuable advancement for autonomous driving and other night-time vision applications.