MemGEN: Memory is All You Need

link: https://arxiv.org/pdf/1803.11203v1.pdf
authors: Sylvain Gelly Karol Kurach Marcin Michalski Xiaohua Zhai
abstract: They have proposed a novel approach for generative modeling based on
Rademacher Coin Flipping metrics
algorithm:
Screen Shot 2018-04-10 at 9.20.59.png
conclusion: comparing to this dataset selection method, this proposed approach can result very well.

Modeling Semantic Plausibility by Injecting World Knowledge

link: https://arxiv.org/pdf/1804.00619v2.pdf
authors: Su Wang, Greg Durrett, Katrin Erk
abstract: They have proposed validated the improvement of accuracy in semantic plausibility by injecting world knowledge to existing neural network. the model is the combination of neural network(3 layers) and world knowledge.
Screen Shot 2018-04-10 at 10.02.25.png
And this is the world knowledge they used in that paper.
Screen Shot 2018-04-10 at 10.02.46.png

Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images

Link: https://arxiv.org/pdf/1804.01654v1.pdf
authors: Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, Yu-Gang Jiang
project page: http://bigvid.fudan.edu.cn/pixel2mesh/
abstract: They have proposed the method to produce the triangular mesh 3D shape from a single colour input image. Example is below.

Screen Shot 2018-04-13 at 10.54.08.png

And the strategy they used in this paper called "Coarse-to-Fine" approach, which is starting with rough shape and then curve the sphere to make it intricate.

Screen Shot 2018-04-13 at 10.57.38.png

As you can see, they have improved the accuracy significantly.

Screen Shot 2018-04-13 at 10.59.35.png

Future work: Our method only produces meshes with the same topology as the initial
mesh. Future work involves extending our approach to handle more general case,
such as scene level reconstruction, and learn from multiple images for multi-view reconstruction.

Sign up for free and join this conversation.
Sign Up
If you already have a Qiita account log in.