rgbd segmentation

Differentiable Physical based Image Enhancement

Introduction:

Computer vision has played an important part in helping agents in understanding and analyzing the real world.
Although computer vision achieved impressive progress in the past few years, a large number of studies have shown
that the performance of many computer vision systems often exhibits a significant drop when they are presented with
common visual turbulences (rain, snow, and fog) and adversarial examples. Existing methods either resort to hand-crafted
priors or using deep learning as a black-box to learn an end-to-end mapping from input to output. However, these priors
can easily violated and black-box model is hard to interpret. On the other hand, we propose to embrace theoretical grounded physical models
with deep learning to make visual sensing more robust. We propose a series of differentiable
physical model to solve low-level vision and high-level understanding problem with state-of-the-art performance

Paper:

• Hongyuan Zhu, Xi Peng, Joey Tianyi Zhou, Songfan Yang, Vijay Chandrasekhar, Liyuan Li, Joo-Hwee Lim
  "RR-GAN: Single Image Rain Removal Without Paired Information",
  The Thirty-thrid AAAI Conference on Artificial Intelligence (AAAI2019, Oral)
  [PDF]
  [Code]
• Hongyuan Zhu, Xi Peng, Vijay Chandrasekha, Liyuan Li, Joo-Hwee Lim
  "DehazeGAN: End-to-End Single Image Dehazing with Generative Adversarial Network",
  Internation Joint Conference of Artificial Intelligence (IJCAI2018, Oral)
  [PDF]
  [Code]

Related Works:

• Xi Peng, Joey Tianyi Zhou, Hongyuan Zhu, "k-meansNet: When k-means Meets Differentiable Programming",
  in arXiv: 1808.07292, 2018. [PDF]
• Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu*, Xi Peng, Yong Liu, Rick Siow Mong Goh. "AnomalyNet: An Anomaly Detection Network for Video Surveillance",
  IEEE Transactions on Information Forensics & Security (TIFS) , 2019.
  [PDF]
  [Code]