SynergyAmodal16K
收藏资源简介:
SynergyAmodal16K是一个高质量的16K样本数据集,基于EntitySeg数据集,涵盖了广泛的实例类别和遮挡场景,并伴有高质量的注释,包括遮挡形状、外观和标题。该数据集旨在解决图像去遮挡(或非模态完成)任务中的数据稀缺问题,通过将自然场景中的图像数据、人类专业知识和生成先验相结合,以实现多样性和逼真度的平衡。数据集的创建过程采用了一种数据-人类-模型协同合成的流水线,首先通过自监督学习算法训练一个部分完成模型,然后通过人类专家的指导和先验模型约束进行筛选和注释,最终生成高质量的配对非模态数据。SynergyAmodal16K数据集适用于AIGC、自动驾驶和机器人等领域,旨在帮助研究人员构建更强大的非模态完成模型,以实现零样本泛化能力和文本可控性。
SynergyAmodal16K is a high-quality 16K-sample dataset based on the EntitySeg dataset. It covers a wide range of instance categories and occlusion scenarios, accompanied by high-quality annotations including occlusion shapes, appearances and captions. This dataset aims to address the data scarcity issue in the task of image amodal completion (or occlusion removal). It balances diversity and realism by combining natural scene image data, human expertise and generative priors. The dataset construction adopts a data-human-model collaborative synthesis pipeline: first, a partial completion model is trained via self-supervised learning algorithms, then filtered and annotated under the guidance of human experts and constraints from prior models, ultimately generating high-quality paired amodal data. The SynergyAmodal16K dataset is applicable to fields such as AIGC, autonomous driving and robotics, and aims to help researchers build more powerful amodal completion models to achieve zero-shot generalization capability and text controllability.
SynergyAmodal 数据集概述
数据集名称
SynergyAmodal 😷⇒☺️
数据集简介
Deocclude Anything with Text Control
数据集状态
Code, Dataset, and Models will be released soon, stay tuned!

- 1SynergyAmodal: Deocclude Anything with Text Control厦门大学 · 2025年



