ROAD-IROS: Automatic Data Annotation Challenge for ROAD Dataset

Co-hosted by IROS 2024

October 15, Abu Dhabi

Competition Agenda

Automating the process of annotations can significantly facilitate the research for autonomous vehicles and increase field-evaluations. Training deep-learning models in the realm of autonomous driving require massive datasets annotated for diverse downstream tasks, such as detections, segmentation, tracking, predictions, interaction features, maneuvers, location marks and etc.
The primary goal of this competition is to design an automated annotation pipeline for the ROAD family datasets, that generates annotations in the existing format along with the textual descriptions. ROAD family datasets are datasets for road events detection. These datasets encompass videos from two different countries: the United Kingdom (ROAD-UK ) and the United States (ROAD-WAYMO). Automatic annotations according to the existing format, will significantly speed up utilization of a newly collected data and reduce the efforts required from human annotators. Textual descriptions provide deeper insights into the images and understanding the environmental context, and can be further utilized by the community for Vision Language Models(VLMs) to train hybrid models for road event detection.

This includes the following tasks:
• Detecting and modeling complex activities, contributed to by several agents over an extended period of time. • Predicting agent intentions. • Forecasting future road events. • Based on all these elements, deciding what action the autonomous vehicle should perform next. • Modelling the reasoning processes of road agents in terms of goals or mental states.

We invite to a challenge participants from both academia and industry (individually and in groups), professionals who are working in the areas of autonomous driving and seeking for automation of the annotation process.

Challenges

Automating the process of annotations can significantly increase field-evaluations in diverse scenarios and environments, while requiring minimum effort from human annotators.
The designed best performing automotive annotation pipeline will be utilized to annotate ROAD-UAE dataset and can be further used for domain adaptation for accurate intention detections. RODA-UAE dataset is an extension of ROAD family datasets created by collecting driving videos from the UAE region. Generating accurate annotations for ROAD-UAE dataset will enable a new research topic for domain adaptation.
Enriching annotations of ROAD-UK and ROAD-WAYMO datasets with textual descriptions, will contribute to the depth and usability of annotated data and support the development of a road event detection model that is not only reactive to direct visual inputs but also proactive in understanding contextual subtleties.