SI-SCORE

SI-SCORE是由谷歌研究大脑团队创建的合成数据集，用于精细分析机器学习模型对物体位置、旋转和尺寸变化的鲁棒性。该数据集包含611,608张图像，通过将614个不同类别的物体粘贴到867个背景上，并系统地改变物体的大小、位置和方向来生成。数据集的创建过程涉及精确控制物体和背景的组合，以及使用OpenImages提供的分割掩码提取前景。SI-SCORE主要用于评估和比较不同类型的深度学习模型，如ResNets、Vision Transformers和CLIP，在处理图像识别任务时的性能和鲁棒性。

SI-SCORE is a synthetic dataset created by the Google Research Brain Team for fine-grained analysis of the robustness of machine learning models against variations in object position, rotation, and scale. This dataset contains 611,608 images, generated by pasting objects from 614 distinct categories onto 867 unique backgrounds while systematically adjusting the objects' scale, position, and orientation. The dataset creation process involves precise control over the pairing of foreground objects and backgrounds, as well as the extraction of foreground regions using segmentation masks provided by OpenImages. SI-SCORE is primarily utilized to evaluate and compare the performance and robustness of diverse deep learning models such as ResNets, Vision Transformers, and CLIP when performing image recognition tasks.