RL agents get specific to tasks they are trained on. What if we remove the task itself during training?

Turns out, a self-supervised planning agent can both explore efficiently & achieve SOTA on test tasks w/ zero or few samples in DMControl from images!

https://ramanans1.github.io/plan2explore

FaceFilter: Audio-visual speech separation using still images

pdf: https://arxiv.org/pdf/2005.07074.pdf
abs: https://arxiv.org/abs/2005.07074
video: https://youtube.com/watch?v=ku9xoLh62E4

Single-Stage Semantic Segmentation from Image Labels

Use save model GAN in Unreal Engine 4

SketchRNN

Neuroevolution in squids experiment

Image Colorization

Project: https://ericsujw.github.io/InstColorization/
Code: https://github.com/ericsujw/InstColorization
Demo: https://colab.research.google.com/github/ericsujw/InstColorization/blob/master/InstColorization.ipynb

TAO: A Large-Scale Benchmark for Tracking Any Object

Images as bar charts of red, green, and blue

Point2Mesh: A Self-Prior for Deformable Meshes

Neural Topological SLAM for Visual Navigation

MediaPipe (https://goo.gle/30hzVAD) has updated its #OpenSource, on-device, real-time hand tracking solution so that it's 15% faster, 37% more accurate and achieves 69% higher recall.

4D Visualization of Dynamic Events from Unconstrained Multi-View Videos