SkyMatte: A High-Quality Dataset for Improving Sky Image Matting
Cheng-Yen Tsai1*
Zheng-Hui Huang1*
Wei-Lien Tang2
Hsiu-Ting Yang2
Jo-Fan Wu2
Yung-Yu Chuang1,3
1National Taiwan University
2ASUSTeK Computer Inc.
3NTU AI-CoRE
*These two authors contributed equally to this work.
Abstract
This paper introduces SkyMatte, a high-quality dataset tailored for sky image matting. The dataset comprises high-resolution images featuring skies under diverse lighting and weather conditions, along with a variety of foreground subjects. The sky mattes are accurate by leveraging a trimap-based image matting model with meticulously annotated trimaps. Experiments demonstrate that the proposed dataset significantly enhances the performance of sky matting models compared to the existing sky matting dataset. Furthermore, we employ a trainable guided filter for sky image matting. The proposed model, trained on the SkyMatte dataset, achieves state-of-the-art performance in sky image matting.
Citation (bibTex)
Cheng-Yen Tsai, Zheng-Hui Huang, Wei-Lien Tang, Hsiu-Ting Yang, Jo-Fan Wu, Yung-Yu Chuang.
SkyMatte: A High-Quality Dataset for Improving Sky Image Matting.
In Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2026),
pages 11362-11366, Barcelona, May 2026.
Paper
ICASSP 2026 paper
Dataset
The SkyMatte dataset (zip file, around 1GB)
Dataset description
The SkyMatte dataset is specifically curated for the sky image matting task. Our dataset contains 746 real-world images captured using mobile devices by ourselves, with a resolution of 2K. For each image, the dataset provides a corresponding alpha matte to delineate the key regions. Out of the 746 images, 726 are allocated to the training set, with the remaining 20 serving as the testing set.
In order to enhance the diversity of the dataset, we aimed to gather a broad spectrum of scenes. This involved capturing various lighting and weather conditions, such as sunny, cloudy, sunset, and night scenes. As a result, the color distribution of the skies varies throughout the dataset. Furthermore, our dataset comprises a diverse array of foreground subjects, including people, buildings, trees, landscapes, and animals, providing a rich semantic range. Additionally, to further enhance and challenge the performance of matting models, we specifically captured some foreground objects with intricate structures, such as hair or leaves.
To ensure the acquisition of accurate mattes as ground truth, we enlisted the assistance of several experts for manual image annotation. However, the process of labeling an alpha matte is labor-intensive, presenting challenges for humans in determining the transparency of the sky for each pixel. To streamline this process, we leveraged trimap-based deep matting models, which enabled experts to annotate trimaps instead. After evaluating various options, we chose to utilize the pre-trained model of FBA Matting as our trimap-based matting network. Annotators had the opportunity to refine trimaps through multiple editing rounds before finalizing the solution. The quality of mattes was further verified by additional experts.
The images are in the standard JPEG format (Joint Photographic Experts Group), while the mattes are in the standard PNG (Portable Network Graphics) format. In the mattes, a value of 255 represents the foreground, while 0 denotes the background (sky). These intensity values are interpreted as alpha values at pixels. Both image and matte files are named with 5-digit indices, such as 00053.jpg. The image and its corresponding matte share the same index but with different file extensions, e.g. 00053.jpg and 00053.png.
Data license
The dataset is provided under CC BY-NC-SA. By accessing and using the dataset, users agree to adhere to the terms and conditions outlined in the data license.
The CC BY-NC-SA license: This license allows reusers to distribute, remix, adapt, and build upon the material in any medium or format for noncommercial purposes only, and only so long as attribution is given to the creator. If you remix, adapt, or build upon the material, you must license the modified material under identical terms.
cyy -a-t- csie.ntu.edu.tw