Unverified50% confidenceFactExact time
ImageNet-C was proposed by Hendrycks et al. in 2019; it applies 15 types of synthetic corruptions to the ImageNet validation set, each with 5 severity levels, forming 75 subsets in total, and its corruptions are image-space post-processing operations rather than being based on real physical imaging processes.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedIn 2019, researchers at the University of Tübingen discovered that CNN models pre-trained on ImageNet rely more on texture features than shape features when recognizing objects, a phenomenon known as texture bias.63% similarUnverifiedFID 由 Heusel 等人于2017年提出，通过比较真实图像与生成图像在 Inception 网络特征空间中的分布距离来衡量生成质量，数值越低越接近真实图像分布62% similarUnverified在深度伪造检测的生产环境中真实图像与伪造图像比例可能高达1000:1甚至更高，极端不平衡下AUC对长尾失败几乎没有区分度62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/977625API
curl https://kongchang.com/api/v1/knowledge/claims/977625MCP
get_claim(id=977625)