E2EAD

A diversity of computer vision capabilities are all critical in building industry-level autonomous driving systems, ranging from 2D to 3D perception, prediction, planning, to scene simulation. This has inspired a surge of relevant research and publications in leading venues such as CVPR, ICCV and ECCV, growing at a fast pace with increasingly accurate and efficient new methods (e.g. BEV-based 3D detection, HDMapNet, NeRF) developed continuously. Much more than simple combination of individual independently developed methods, autonomous driving also requires synergistic integration of different functions as a whole. This however is far away from the current situation that researchers in the sub-fields of perception, planning and simulation make largely limited idea exchange and communication. This calls for a system-level perspective on the advancement of autonomous driving, e.g. to establish end-to-end realtime AI systems streamlining multiple sensors (camera, LiDAR, Radar) and all the necessary functions/tasks (e.g. object detection, lane detection, HD map construction, path planning) under a single umbrella. To achieve this accomplishment, we are still faced with fundamental challenges such as (1) out-of-date visual perception and path planning methods, (2) overly simple integration strategies, (3) lack of high-quality simulation environments. In this context, this workshop aims to provide a platform where researchers from different sub-fields can focus on exchanging the frontier ideas across boundaries, leading to holistic system-aware understanding and systematic research attempts in the future.

• 1. 3D obejct detetion
• 2. Traffic lane detection and HD map construction
• 3. End-to-end perception and path planning
• 4. Autonomous driving environment simulation

Specifically, with Topic 1 and Topic 2, we encourage the researchers to illustrate the frontier of visual perception for autopilot. Whilst the researchers working on Topic 3 could share their thinking in path planning from the end-to-end perception viewpoint. The Topic 4 investigates the training data issue involved in all the previous topics by real-world rendering and simulation. Other topics related to end-to-end driving system design are all welcome and valuable to discuss.