More interesting news on E2E progress.
TL;DR: SenseTime got Best Paper Award at CVPR2023 for a paper that proposes a new groundbreaking end-to-end model for autonomous driving. The model called UniAD (Unified Autonomous driving) combines full stack driving tasks, including detection, mapping in perception, occupancy forecasts, and planning into a single NN. It uses transformer decoder structure with task queries as interfaces connecting each node.
More details:
SenseTime, a leading global artificial intelligence (AI) software company, and its joint labs were recognized for its AI research at the Computer Vision and Pattern Recognition Conference (CVPR) 2023, the world’s largest annual conference on computer vision and AI. Two papers from SenseTime were included in the list of 12 award candidate papers. Among them, the research paper “Planning-oriented Autonomous Driving” won the Best Paper Award.
The winning paper, “Planning-oriented Autonomous Driving” introduces Unified Autonomous Driving (UniAD), a comprehensive up-to-date framework that incorporates full-stack driving tasks, including detection, tracking, mapping in perception, as well as motion and occupancy forecast in prediction in one network.
Instead of a simple stack of tasks, UniAD investigates the effect of each module in perception and prediction, leveraging the benefits of joint optimization from preceding nodes to final planning in the driving scene. All perception and prediction modules are designed in a transformer decoder structure, with task queries as interfaces connecting each node. This innovative approach is a significant breakthrough in autonomous driving technology research and was recognized with the first CVPR best paper award for this field.
Prof. Wang Xiaogang, Co-founder, Chief Scientist and President of Intelligent Automotive Group at SenseTime, said, “UniAD is the industry's first end-to-end autonomous driving foundation model that unites all nodes from perception and decision-making as a whole, leading to significant improvements in overall system performance and representing the forthcoming development trend of autonomous driving.”
With extensive ablations, the effectiveness of using the UniAD is proven by substantially outperforming previous state-of-the-arts (SOTA) in all aspects. For example, multi-object tracking accuracy exceeds SOTA by 20%, online mapping accuracy in lanes is increased by 30%, motion forecasting error is decreased by 38%, and planning error is reduced by 28%.
Source: SenseTime Wins the Best Paper Award at CVPR 2023 For Groundbreaking Perception-Decision Integrated Autonomous Driving Foundation Model-Newsroom-SenseTime
TL;DR: SenseTime got Best Paper Award at CVPR2023 for a paper that proposes a new groundbreaking end-to-end model for autonomous driving. The model called UniAD (Unified Autonomous driving) combines full stack driving tasks, including detection, mapping in perception, occupancy forecasts, and planning into a single NN. It uses transformer decoder structure with task queries as interfaces connecting each node.
More details:
SenseTime, a leading global artificial intelligence (AI) software company, and its joint labs were recognized for its AI research at the Computer Vision and Pattern Recognition Conference (CVPR) 2023, the world’s largest annual conference on computer vision and AI. Two papers from SenseTime were included in the list of 12 award candidate papers. Among them, the research paper “Planning-oriented Autonomous Driving” won the Best Paper Award.
The winning paper, “Planning-oriented Autonomous Driving” introduces Unified Autonomous Driving (UniAD), a comprehensive up-to-date framework that incorporates full-stack driving tasks, including detection, tracking, mapping in perception, as well as motion and occupancy forecast in prediction in one network.
Instead of a simple stack of tasks, UniAD investigates the effect of each module in perception and prediction, leveraging the benefits of joint optimization from preceding nodes to final planning in the driving scene. All perception and prediction modules are designed in a transformer decoder structure, with task queries as interfaces connecting each node. This innovative approach is a significant breakthrough in autonomous driving technology research and was recognized with the first CVPR best paper award for this field.
Prof. Wang Xiaogang, Co-founder, Chief Scientist and President of Intelligent Automotive Group at SenseTime, said, “UniAD is the industry's first end-to-end autonomous driving foundation model that unites all nodes from perception and decision-making as a whole, leading to significant improvements in overall system performance and representing the forthcoming development trend of autonomous driving.”
With extensive ablations, the effectiveness of using the UniAD is proven by substantially outperforming previous state-of-the-arts (SOTA) in all aspects. For example, multi-object tracking accuracy exceeds SOTA by 20%, online mapping accuracy in lanes is increased by 30%, motion forecasting error is decreased by 38%, and planning error is reduced by 28%.
Source: SenseTime Wins the Best Paper Award at CVPR 2023 For Groundbreaking Perception-Decision Integrated Autonomous Driving Foundation Model-Newsroom-SenseTime
