A Multi-Purpose Realistic Haze Benchmark With Quantifiable Haze Levels and Ground Truth

Area of Science:

• Computer Vision
• Image Processing
• Remote Sensing

Background:

• Outdoor imagery is often degraded by smoke or haze, impacting scene understanding.
• Existing datasets lack realism and comprehensive data for evaluating computer vision in degraded visual environments (DVE).

Purpose of the Study:

• Introduce the first realistic haze image benchmark dataset.
• Provide paired haze-free and hazy images with in-situ haze density measurements.
• Facilitate evaluation of computer vision algorithms in DVE.

Main Methods:

• Generated realistic haze using professional smoke machines in a controlled environment.
• Captured images from both unmanned aerial vehicle (UAV) and unmanned ground vehicle (UGV) perspectives.
• Collected paired haze-free images and in-situ haze density measurements.

Main Results:

• Developed a novel benchmark dataset for DVE research.
• Dataset includes ground truth object classification bounding boxes and haze density.
• Evaluated state-of-the-art dehazing and object detection algorithms on the dataset.

Conclusions:

• The introduced dataset addresses the need for realistic benchmarks in DVE research.
• The dataset enables robust evaluation of computer vision algorithms under haze conditions.
• The dataset is publicly available for community use and algorithm development.