Skip to content

Latest commit

 

History

History
17 lines (10 loc) · 1017 Bytes

objectformer.md

File metadata and controls

17 lines (10 loc) · 1017 Bytes
description
ObjectFormer for Image Manipulation Detection and Localization

ObjectFormer

J. Wang et al., “ObjectFormer for Image Manipulation Detection and Localization,” 2022 Ieee Cvf Conf Comput Vis Pattern Recognit Cvpr, vol. 00, pp. 2354–2363, 2022, doi: 10.1109/cvpr52688.2022.00240.

作者提出了使用frequency特征以及将它和RGB特征结合,以embedding的形式输入到encoder里面,主要的贡献点是1. 用了这个high frequency feature,2. 对frequency feature应用了一个边界敏感上下文不一致模块boundary-sensitive contextual incoherence module,用来对特征做fine-grain,3. 利用了object prototype,学习mid-level feature。

作者在做消融研究的时候,baseline是efficientNet-B4?因为他用的backbone就是efficient-B4,

Result

image-20230711173825281