Will do a write-up later.
#CVPR20 #computervision
1. Dynamic Graph Message Passing Networks arxiv.org/pdf/1908.06955…
It addresses the modelling long-range dependencies problem by using feature map as a feature vector nodes and dynamically sample the neighborhood of a node from the feature graph.
A generative image model that can leverage the feature space from different semantic levels learned by a pretrained classification network. many generative applications to play with
Inspired by idea of Bert in #nlp, the authors propose to contruct a large and consistent dictionnaries for unsupervised learning with constrative loss.
I found it quite similar representation learning idea as Electra in #nlp. It proposes Limited Context Inpainting by masking regions of image and disciminate different transformations
![Image](https://pbs.twimg.com/media/EayiKCeWAAAuvlm.jpg)
The authors challenge the common belief in knowledge distillation by showing that the student can also enhance the teacher significantly and poorly trained-teacher can help the student
#FAIR researchers did an amazing job on human 3D model from a single image! impressive results!
A paper on Model efficiency of object detection by introducing weighted bi-directional feature pyramid network (BiFPN) for efficient multi-scale feature fusion and a novel compound scaling method
![Image](https://pbs.twimg.com/media/Ea_CpnpWkAAKbcz.png)
The authors propose an autoencoder which can generate images with quality comparable to state-of-the-art GANs while also learning a less entangled representation.
Some impressive results with reconstruction of images:
![Image](https://pbs.twimg.com/media/Ea_D1TtXQAAk9ao.jpg)