ImageNet-D

ImageNet-D是由韩国科学技术院创建的一个合成图像数据集，通过扩散模型生成，包含4835张图像，涵盖113个类别和547种不同的背景、纹理和材料。该数据集用于评估和提高深度学习模型在视觉识别任务中的鲁棒性，特别是在面对复杂和多样化的视觉变化时的表现。ImageNet-D通过引入高级别的图像变化因素，如背景、纹理和材料的多样性，挑战现有模型，并推动模型在未见过的复杂场景中的性能提升。

ImageNet-D is a synthetic image dataset developed by the Korea Advanced Institute of Science and Technology, generated via diffusion models. It comprises 4,835 images spanning 113 categories and 547 distinct backgrounds, textures, and materials. This dataset is designed to evaluate and enhance the robustness of deep learning models in visual recognition tasks, particularly their performance against complex and diverse visual variations. By introducing high-level image variation factors such as the diversity of backgrounds, textures, and materials, ImageNet-D challenges existing models and promotes improvements in model performance on unseen complex scenarios.