Seminar

CVL Talk: Towards Reducing Human Efforts in Visual Perception and Beyond - Kaicheng Yu, Alibaba Group

Abstract:

Although GPT-based AI models have achieved widespread popularity, their virtual nature restricts their ability to reduce human effort in physical tasks, highlighting the importance of robots in realizing genuine AI democratization. At Alibaba, we have deployed over 700 autonomous vehicles for delivery, and we have identified that there are two major challenges preventing further scaling up. Firstly, expert intervention is necessary to maintain perception modules that contain neural networks. To this end, we made several progresses in minimizing human effort through neural architecture search. Additionally, we recognize the need for 3D data annotation proportional to deployment scale. Our solution is to develop an autolabing system with a simple yet robust framework that fundamentally tackles the failure of multi-sensors setup. In closing, we present our latest effort, AutoML Perception V1, which is a broad-based autoML AI system designed to address both of the aforementioned challenges in an industrial setting. Deployed on a cloud platform, this system reduces human effort on average by 95% while significantly outperforming baseline models and drastically lowering costs.

Bio:

Dr. Kaicheng Yu is a senior research scientist in Alibaba Group. His research interests cover broader area of computer vision, machine learning, and artificial intelligence, and specifically building an intelligent visual system. He receives Westlake Star Award (Young Talent Program) at 2022, Qualcomm Innovation Fellowship Europe at 2019, and has been recognized as outstanding reviewer at ECCV2020 and CVPR2021. He publishes on top conference venues like CVPR, ICCV, ECCV, ICML, ICLR and NeurIPS, and journals like TPAMI.

Zoom Link

You can see the other CVPR 2023 UBC events here.