主要收集從 2016年11月(CVPR2017 deadline)到現在的 生成對抗網路(GAN)相關paper (按arXiv發表順序), 有遺漏歡迎補充。

[1611。04076] Least Squares Generative Adversarial Networks (Cycle GAN的D用了其中的方法,將Loss改為L2 Loss,訓練穩定性提高了,好於傳統的cross-entropy loss)

[1611。07004] Image-to-Image Translation with Conditional Adversarial Networks (pix2pix) 這篇中70x70的patchD在cyclegan中也有延續。 注:這篇中是pair來訓練的

[1612。05363] Learning Residual Images for Face Attribute Manipulation (CVPR 2017) 這一篇比較早就用dual learning了,比cycle gan都早了3個月啊。但是他用的cycle loss是D的Loss,而不是pixel level的L1 Loss。

[1612。07828] Learning from Simulated and Unsupervised Images through Adversarial Training (CVPR 2017)

[1701。07717] Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro (ICCV2017) 這篇索性拿G產生的圖片來做資料增強,做了一個半監督學習框架,也容易理解。訓傳統分類ResNet,而沒有使用D。 因為D現在還是比較淺?在Fine-grain dataset上也有提高。

[1701。02676] Unsupervised Image-to-Image Translation with Generative Adversarial Networks 之前我們可以用cgan來指定生成什麼domain。作者多加了一個分類器來預測隨機input z(在學習對映完成後,z其實是有semantic含義的)所以如果除了c以外,我們還能指定z。

[1701。07875] Wasserstein GAN

[1703。02291] Triple Generative Adversarial Nets

[1703。05192] DiscoGAN: Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (ICML2017) 同一個世界,同一個夢想 喜歡這篇裡面的圖和實驗。

[1703。10593] CycleGAN: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (ICCV2017) 同一個世界,同一個夢想 程式碼很solid。 D用的是L2 Loss, Cycle Loss和Identity Loss都是L1 Loss。

[1704。02510] DualGAN:Unsupervised Dual Learning for Image-to-Image Translation (ICCV2017) 同一個世界,同一個夢想

[1704。00028] Improved Training of Wasserstein GANs

Visual Saliency Prediction with Generative Adversarial Networks 阿嶽感覺很一般

Boosting Generative Models

Towards Realistic High-Resolution Image Blending

[1704。05838] Generative Face Completion

[1704。04086] Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis (ICCV 2017) 人臉pose旋轉,需要pair來訓練。一個網路做global的人臉,一個網路切5個關鍵點。最後融合到一起。

[1704。04131] Neural Face Editing with Intrinsic Image Disentangling (CVPR2017)

[1706。05274v2] Perceptual Generative Adversarial Networks for Small Object Detection

Decomposing Motion and Content for Video Generation

[1706。07068]Creative Adversarial Networks, Generating “Art” by Learning About Styles and Deviating from Style Norms 生成藝術作品

[1707。03124] Adversarial Generation of Training Examples for Vehicle License Plate Recognition 用GAN來生成車牌,做資料增強

我的其他文章:

用GAN生成的影象做訓練?Yes!

2017 ICCV 對抗生成網路GAN接收論文

2017 ICCV 行人檢索/重識別 接受論文彙總