Jonathan Long

I am a researcher/engineer at Symbio Robotics, bringing modern perception to industrial robots.

Before that, I completed a PhD at UC Berkeley, working on computer vision and advised by Trevor Darrell.

Before 2010, I was at Carnegie Mellon, studying Computer Science, Physics, and Mathematics.

I care about recognition: algorithms that learn to decode perceptual input into useful information.

We've figured out how to build powerful supervised recognition machines, and now it's time to bring that technology to manipulation of the physical world.

I obsess about having the most powerful tools for research and engineering; I'm a developer of Caffe.

jon@symb.io

@longjon on GitHub

Fully Convolutional Networks for Semantic Segmentation

Jon Long*, Evan Shelhamer*, Trevor Darrell (CVPR 2015 best paper honorable mention)
*equal contribution

Fully convolutional networks by themselves, trained end-to-end on segmentation data, initialized from recent classification models, and with extra links between nonconsecutive layers, improve semantic segmentation on PASCAL by 20% relative.

Do Convnets Learn Correspondence?

Jon Long, Ning Zhang, Trevor Darrell (NIPS 2014)

Max-pooling convolutional networks trained on classification perform surprisingly well at fine-scale tasks like alignment and keypoint prediction.