From Graphics, to Vision, to the Omnipresence of Deep Learning: Adapting to Making the Most Impact

This talk is probably different from a typical academic talk, in which I will give a broad introduction about my immersive experience in both academic and industry world.

I will start by introducing my early work on 3D reconstruction and display, aiming to bring distant people to collaborate over space as if they were in the same room. The line of research slowly moved to the more challenging and open problem of complete 3D reconstruction from incomplete and noisy input, we found that by introducing semantics can significantly improve the reconstruction quality. Then I will present works done at Baidu Research on autonomous driving, from dataset curation, AR simulation, learning based navigation scheme, to making an excavator autonomously digging trenches. Finally I will talk about the commercialization of autonomous driving on truck, its unique technical challenges and how we address them in a production setting.

Looking forward, in the era of big model (with big expense), I will present a few possible directions to bridge the gap between academic research, industrial research and commercialization.