Research Blog from OpenDriveLab at CVPR 2023
Join the world’s top minds in AI at IEEE/CVF CVPR 2023! OpenDriveLab is excited to be an active part of this prestigious event, taking place in Vancouver from June 18, 2023. Explore the future of autonomous driving with the cutting-edge research showcased in our latest work!
This year, our focus is on end-to-end autonomous driving, bird’s-eye-view (BEV) perception, and foundation model, resulting in 1 Best Paper Award Candidate, 1 Highlight Paper, 3 Posters, 2 Workshop Events and 1 Challenge.
Unfortunately, most of us cannot go to Vancouver in person due to the insane visa issue. We do have two teammates onsite, please say hello to Yunsong and Tutian! :)
Best Paper Award Candidate: Planning-oriented Autonomous Driving
We present UniAD (Unified Autonomous Driving), a planning-oriented end-to-end framework for autonomous driving. We wish this could initiate a new step of foundation model for autonomous driving and robotics.
Highlight Paper: BEVFormer V2
Notably, BEVFormer V1 is selected as one of ECCV 2022 Top-10 Influential Papers and the 100 most cited AI papers in 2022.
This year, we present BEVFormer V2, a novel BEV detector with perspective supervision. It converges faster and better suits modern image backbones such as InternImage.
- Poster venue: THU-AM-129
- Third-party blog in Chinese is here.
Posters
MaskAlign. Is reconstruction necessary in Masked Image Modeling (MIM) with a teacher model? We propose a much more efficient paradigm named MaskAlign without any reconstruction on masked tokens, and apply the alignment on visible patches only. MaskAlign achieves new state-of-the-art performance with much higher efficiency.
- Code and model checkpoints is released at Github here.
- Poster venue: THU-PM-203
- Third-party blog in Chinese is here.
ThinkTwice is an end-to-end autonomous driving framework with the focus on scaling up the decoder part.
- Code and models are hosted at Github here.
- Poster venue: THU-PM-131
- Third-party blog in Chinese is here.
FocalDistiller is a knowledge distillation algorithm designed for extracting focal knowledge from an imperfect bird’s-eye-view (BEV) expert.
- Poster venue: TUE-AM-094
Workshop and Challenge in Two Cities
This year we present End-to-end Autonomous Driving workshop, aiming to bring together leading researchers from both academia (UCB, CMU) and industry (Tesla, NVIDIA, Cruise, Wayve, XPENG), to shed new insight on pipelines / paradigms / frameworks for autonomous vehicles. Check out the amazing speakers along with their talk videos.
Along with the workshop came up with the Autonomous Driving Challenge, featuring a total cash award of USD $100,000 and four all-brand-new tasks.
We received more than 2300 submissions from over 270 teams, covering over 15 countries and regions. Announcement of the Challenge winners is posted in a previous blog. The Challenge wrapped up with total success. Thanks everyone!
Since most Chinese participants cannot make it to CVPR, we do actually prepare a sibling event with awesome speakers at Beijing on June 18. Check out the fantastic program here!
Professional Service
Congratulations to Ilnehc and Chonghao Sima for being recognized as Outstanding Reviewers! This is a small group of professionals (200+) out of a large pool of reviewers over 7000. The community is stronger with their aids! Besides, many researchers / PIs at Shanghai AI Lab serve as (senior) area chairs for CVPR, altogether we are building a more professional community to shape the future of AI.
That’s all from OpenDriveLab. Please, enjoy CVPR and get to know people! Last, we are always at your service via contact@opendrivelab.com