InstanceAugmentation

InstanceAugmentation是由牛津大学视觉几何组创建的数据集，专注于通过实例级增强来扩展图像数据集。该数据集通过重新绘制图像中的单个对象，保持其原始形状，从而在不改变类别标签的情况下增加数据多样性。数据集创建过程中，结合了预训练的潜在扩散模型，通过深度和边缘图控制条件，无缝重新绘制场景中的个体对象。此方法不仅适用于任何分割或检测数据集，还特别适用于数据匿名化，如重新绘制隐私敏感实例。InstanceAugmentation的应用领域包括对象检测、语义分割和显著性分割，旨在通过增加数据集的视觉变异性来提高模型的性能和泛化能力。InstanceAugmentation is a dataset developed by the Visual Geometry Group at the University of Oxford, which focuses on expanding image datasets via instance-level augmentation. It increases data diversity without altering category labels by redrawing individual objects in images while preserving their original shapes. During the dataset creation process, a pre-trained latent diffusion model is integrated, with depth and edge maps used as control conditions to seamlessly redraw individual objects in the scene. This method is applicable to any segmentation or detection dataset, and is particularly suitable for data anonymization tasks such as redrawing privacy-sensitive instances. The application fields of InstanceAugmentation include object detection, semantic segmentation and saliency segmentation, aiming to improve model performance and generalization ability by increasing the visual variability of datasets.