三个分支:擦除、转换、噪声 用来生成对抗性表征,模拟遮挡问题 对应信息丢失、位置错位和噪声信息
对抗性防御:思路是GAN网络,以对抗性的方式优化生成器和判别器
- Re-Identification Identification Transforming Occluded Erasingre-identification identification transforming re-identification re-identification identification information re-identification identification co-parsing re-identification identity-guided identification re-identification intra-identity identification camera camera-conditioned re-identification re-identification identification perspectives re-identification identification adaptive pairwise occluded