CARLA-Haze: A Synthetic Benchmark for Outdoor Image Dehazing

Henry Velesaca, Leo Thomas Ramos, Ángel D. Sappa,
ESPOL Polytechnic University, Ecuador
Computer Vision Center, Universitat Autònoma de Barcelona, Spain
Paper GitHub Data

CARLA-Haze has been accepted at the 5th IEEE/CVF Workshop on Image/Video/Audio Quality Assessment in Computer Vision, VLM and Diffusion Model (WVAQ) @ WACV 2026 📣📣📣

CARLA-Haze contains 10,000 high-resolution paired images (clean and hazy), distributed across 10 different scenarios and 10 incremental haze intensity levels. The dataset includes diverse scenarios and visual elements, distance-based haze distribution consistent with realistic atmospheric conditions, and multiple viewpoints enhancing visual variability. Additionally, two preprocessed versions with standardized resolutions of 640x480 and 512x512 pixels are provided, each containing 40,000 images. It also offers predefined splits for training, validation, and testing to facilitate its use.

Abstract

This project presents CARLA-Haze, a synthetic dataset designed for outdoor image dehazing. CARLA-Haze contains 10,000 high-resolution paired images (clean and hazy), distributed across 10 different scenarios and 10 incremental haze intensity levels. The dataset includes diverse scenarios and visual elements, distance-based haze distribution consistent with realistic atmospheric conditions, and multiple viewpoints enhancing visual variability. Additionally, two preprocessed versions with standardized resolutions of 640×480 and 512×512 pixels are provided, each containing 40,000 images. CARLA-Haze also offers predefined splits for training, validation, and testing to facilitate its use.

Main features of CARLA-Haze

10 haze intensity levels

CARLA-Haze contains 10 haze intensity levels. These range progressively from 10% (light haze) up to 100% (dense haze), providing a comprehensive set of scenarios that simulate varying visibility conditions.

Distance-based haze distribution

CARLA-Haze incorporates distance-based atmospheric scattering according to CARLA's physically-consistent haze model. In this model, distant objects are more affected by haze, appearing more occluded due to increased scattering, while closer objects maintain higher visual clarity.

Diverse scene elements

CARLA-Haze contains a diverse set of visual elements, including buildings, vehicles, vegetation, street infrastructure, and recreational objects, providing high visual variability.

Visual perspective variability

CARLA-Haze contains images captured at different viewpoints and distances within the same scenario, resulting in variations of visual perspective. Consequently, scene elements exhibit positional and scale changes, enhancing the dataset's visual richness.

BibTeX

If you find this dataset useful, please cite our paper: 
@InProceedings{carla_haze_2026_WACV,
    author    = {Velesaca, Henry O. and Ramos, Leo Thomas and Sappa, Angel D.},
    title     = {CARLA-Haze: A Synthetic Benchmark for Outdoor Image Dehazing},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
    year      = {2026},
    pages     = {1-10}}