RevealLayer-100K
收藏资源简介:
RevealLayer Open数据集是一个用于图像分层分解的开源数据集,源自ICML 2026论文《RevealLayer: Disentangling Hidden and Visible Layers via Occlusion-Aware Image Decomposition》。该数据集旨在支持基于边界框引导的自然图像分解任务,即给定一张RGB图像和实例边界框,将场景分解为干净的背景层和对象级的前景层(每个前景层以RGBA格式表示)。数据集规模在10万到100万样本之间。数据内容包含两个主要部分:1) 训练集和完全标注的基准子集(RevealLayerBenchMark-200),提供原始图像、背景图像真值、前景RGBA层真值以及边界框标注;2) 野外基准子集(RevealLayerBenchMark-wild),仅包含原始图像和边界框标注,用于定性评估和真实世界鲁棒性测试,不提供背景或前景层真值。每个样本的元数据通过JSON文件组织,包含唯一标识符、图像路径、背景路径、前景层路径列表和边界框列表等字段。该数据集适用于图像分解、分层图像编辑、对象移除、图像抠图、图像修复等计算机视觉任务。需要注意的是,数据集不包含第三方基准(如AIM-500、RefMatte_RW100、OBER-Test/ObjectClear)的图像和真值,用户需从官方来源获取并遵守相应许可。
The RevealLayer Open dataset is an open-source dataset for image layer decomposition, originating from the ICML 2026 paper RevealLayer: Disentangling Hidden and Visible Layers via Occlusion-Aware Image Decomposition. It aims to support bounding box-guided natural image decomposition tasks, where given an RGB image and instance bounding boxes, scenes are decomposed into a clean background layer and object-level foreground layers (each represented in RGBA format). The dataset size ranges from 100,000 to 1,000,000 samples. It consists of two main parts: 1) a training set and a fully annotated benchmark subset (RevealLayerBenchMark-200), providing original images, background ground truth, foreground RGBA layer ground truth, and bounding box annotations; 2) a wild benchmark subset (RevealLayerBenchMark-wild), containing only original images and bounding box annotations for qualitative evaluation and real-world robustness testing, without background or foreground layer ground truth. Metadata for each sample is organized via JSON files, including fields such as unique identifiers, image paths, background paths, lists of foreground layer paths, and bounding box lists. The dataset is suitable for computer vision tasks like image decomposition, layered image editing, object removal, image matting, and image inpainting. Note that the dataset does not include images and ground truth from third-party benchmarks (e.g., AIM-500, RefMatte_RW100, OBER-Test/ObjectClear), and users must obtain them from official sources and comply with respective licenses.





