Join JingDong Wang - a Principal Research Manager from MSR Asia on a discussion of HRNet; the most exciting and powerful new vision CNN architecture to come out since AlexNet! This is seriously impressive and makes me incredibly proud to work for Microsoft. Just like in 2015 when we produced the best-of-breed vision architecture RESNET – this takes it to a whole new level.

HRNET is home-grown at Microsoft and has already demonstrated state-of-the-art performance in image classification, semantic segmentation, object detection, facial landmark detection and human pose estimation! Actually; this architecture would probably be state-of-the-art in almost every vision task – they just haven’t had time to implement on all upstream tasks yet!

The idea with HRNET is that it maintains a high resolution path throughout the entire network with cross-resolution pathways achieved with carefully designed “fusions”. Most traditional CNN architectures were designed in the days of image classification where high resolution was not important.

High resolution encoder networks now like Hourglass, Segnet, UNET etc, recover the high resolution from low resolution losing a lot of vital information in the process.

Previous CNN architectures modelled multi-resolution in series;

JingDong will;

Discuss the evolution of vision architectures leading up to HRNETDiscuss ablation studyDiscuss how to use HRNET for different upstream tasks and the respective SOTA performance and resultsPractical guidance on how to use HRNET

Check out HRNet on Github -- https://github.com/HRNet Paper link: https://arxiv.org/abs/1904.04514

The call will be chaired by Karol Zak and Tim Scarfe

Incredibly excited about this one! Go Microsoft!