InsCore (Instance Core Segmentation Dataset)
收藏资源简介:
InsCore是一个用于工业实例分割的合成预训练数据集。该数据集基于公式驱动的监督学习(FDSL)生成,能够生成完全标注的实例分割图像,反映工业数据的特征,包括复杂的遮挡、密集的分层掩膜和多样的非刚性形状。与传统的真实图像数据集相比,InsCore不依赖于真实图像或人工标注,且在五个工业数据集上的实验表明,使用InsCore预训练的模型在实例分割性能上超过了在COCO和ImageNet-21k上训练的模型,以及微调的SAM模型,平均提高了6.2个点的性能。
InsCore is a synthetic pre-training dataset for industrial instance segmentation. It is generated based on Formula-Driven Supervised Learning (FDSL), which can produce fully annotated instance segmentation images that reflect the core characteristics of industrial data, including complex occlusions, dense layered masks and diverse non-rigid shapes. Compared with traditional real-world image datasets, InsCore does not rely on real images or manual annotations. Experiments conducted on five industrial datasets demonstrate that models pre-trained with InsCore outperform those trained on COCO and ImageNet-21k, as well as fine-tuned SAM models, with an average performance improvement of 6.2 percentage points.




