Real-World Video Datasets for Haze Removal
Description
Description for the Video Dehazing Dataset The proposed dataset comprises 22 synthetic hazy videos, each carefully designed to simulate varying levels of haze intensity and diverse environmental conditions, such as different lighting, weather patterns, and scene complexities. Each video was converted into individual frames, resulting in a comprehensive dataset suitable for training and evaluating advanced video dehazing algorithms. This dataset was utilized to explore a novel machine-learning-based approach that leverages the UNet architecture in conjunction with a linear variance scheduler within the diffusion process framework. The frames serve as input to the dehazing model, enabling the system to learn spatiotemporal features effectively. The dataset provides a valuable benchmark for researchers focusing on video dehazing and restoration tasks, offering high-quality synthesized data to test innovative techniques in image enhancement and haze removal. It is particularly suited for algorithms requiring extensive frame-by-frame processing while preserving temporal consistency.
Files
Steps to reproduce
The dataset was independently created to support research in video dehazing, featuring 22 hazy videos generated without corresponding ground truth (clean) videos. The following steps describe how the data was collected, prepared, and processed, enabling other researchers to replicate the workflow: Video Acquisition: The hazy videos were captured or sourced from synthetic environments designed to simulate real-world hazy conditions. Simulations incorporated varying levels of haze intensity, scene complexity, and motion dynamics to ensure diverse data representation. Frame Extraction: Each video was converted into individual frames using tools such as FFmpeg. This process resulted in a high-quality collection of hazy images suitable for frame-based dehazing model training and testing. Preprocessing: The extracted frames were resized to standard dimensions for consistency and normalized to improve compatibility with machine learning models. Basic metadata, such as frame count and video resolution, was documented for future reference. Application of UNet and Diffusion Processes: The dataset was used to develop and evaluate a novel dehazing method combining the UNet architecture with a linear variance scheduler within a diffusion process framework. This approach focuses on learning haze removal directly from hazy inputs without relying on paired ground truth data. Evaluation Metrics: Since no ground truth videos are available, performance evaluation was conducted using non-reference image quality metrics such as PSNR(Peak Signal-to-Noise Ratio) and SSIM(Structural Similarity Index Measure). These metrics helped assess the effectiveness of the dehazing process in improving visual quality. Workflow Transparency: The entire data creation and processing workflow were meticulously documented, including the methods and tools used. Open-source Python libraries like PyTorch, OpenCV, and NumPy were employed to ensure accessibility for other researchers. Dataset Validation: Each video and corresponding frames were manually reviewed to verify the quality of the hazy effects and eliminate inconsistencies. This step ensures the dataset is robust and suitable for training and evaluating video dehazing algorithms.