GAN¶
- General Enhancement
- Training Stability
- Conditional / Repersentation of GAN
- image-to-image GAN
- cycleGAN
- UNIT
- GAN: improve resolution
- GAN for data augmentation
- Adversarial Domain Adaptation
- Person-reID GAN (ICCV 2017)
- BAGAN
- DeLiGAN (CVPR 2017)
- VERI-Wild (CVPR 2019)
- FD-GAN: Pose-guided feature distilling GAN for robust person re-identification (NIPS 2018)
- Multi-pseudo regularized label for generated samples in person re-identification. (TIP 2018)
- Pose transferrable person re-identification (CVPR 2018)
- Pose-normalized image generation for person re-identification (ECCV 2018)
- DG-Net (2019)
- Recover and identify: A generative dual model for cross-resolution person re-identification
- Adversarial Attacks
Challenges¶
Mode Collapse / Mode Diversity¶
generator collapses, produces limited varieties of sample to fake discrimator cuased by: dimension deficient solutions: generalized dataset, Orthonormal constraint (正交約束), diversity loss (e.g. multi-scale structural similarity (MS-SSIM))
Non-convergence, Convergence difficulty¶
the model parameters oscillate, destabilize
Vanishing gradient because of too good discriminator¶
the discriminator gets too successful that the generator gradient vanishes and learns nothing solution: WGAN - Wasserstein distance
Image quality¶
- resolution
- noise
- high frequencey