Night-Time Scene Parsing With a Large Real Dataset.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Scene analysis typically focuses on daytime images, neglecting challenges in night-time conditions.
• Existing methods lack effective handling of diverse lighting issues like over- and under-exposure in night images.
• Labeled night-time datasets for semantic segmentation are scarce, hindering research progress.

Purpose of the Study:

• To address the limitations of current night-time scene parsing (NTSP) techniques.
• To create a comprehensive dataset for training and evaluating NTSP models.
• To develop an advanced framework capable of handling complex exposure variations in night-time imagery.

Main Methods:

• Collected and annotated NightCity, a novel dataset comprising 4,297 real night-time images with pixel-level semantic annotations.
• Proposed an exposure-aware framework that integrates learned exposure features into the segmentation process.
• Conducted extensive experiments to validate the dataset and the proposed framework's effectiveness.

Main Results:

• Training on the NightCity dataset significantly enhances NTSP performance.
• The proposed exposure-aware model achieves state-of-the-art results on multiple benchmarks.
• The framework demonstrates superior performance in parsing complex night-time scenes with varying exposures.

Conclusions:

• The NightCity dataset is a valuable resource for advancing night-time scene parsing research.
• The exposure-aware framework offers a robust solution for semantic segmentation in challenging low-light environments.
• This work sets a new benchmark for performance in night-time scene understanding.