29 Jul SETLabs at the IEEE IV 2024 in Korea
Intelligent Vehicle Symposium
As a premier conference sponsored by the IEEE Intelligent Transportation Systems Society (ITSS), the 35th Intelligent Vehicles Symposium once again stood as a beacon of excellence in intelligent vehicles and systems. We are proud that this year two representatives from our SETLabs team were invited by the organizers to present their posters.
The 35th IEEE Intelligent Vehicle Symposium, held in the picturesque setting of Jeju Island, Korea, brought together a diverse group of researchers, engineers, and practitioners from industry and academia. The symposium served as a platform for presenting the latest advancements in intelligent vehicles and systems.
Among the participants were Max Ronecker and Ahmed Ghita, two PhD candidates from SETLabs Research, who were invited to showcase their latest research during the poster sessions.
Max Ronecker introduced a groundbreaking hybrid architecture designed to enhance radar-based Dynamic Occupancy Grid Mapping (DOGM) for autonomous vehicles. His innovative approach integrates deep learning for state classification, significantly improving the accuracy and reliability of DOGM systems.
Ahmed Ghita presented ActiveAnno3D, the first active learning framework for multi-modal 3D object detection. This framework promises to advance the field of autonomous vehicle perception by optimizing the annotation process and improving detection performance across multiple sensor modalities.
The symposium provided a valuable opportunity for Max and Ahmed to engage with leading experts, exchange ideas, and receive feedback on their research, contributing to the ongoing development of intelligent vehicle technologies.
Dynamic Occupancy Grid Mapping
How to make it work with Radar?
As highlighted at ICRA 2024, radar-based Dynamic Occupancy Grid Mapping (DOGM) for autonomous vehicles is effective in object detection. However, it sometimes faces difficulties in distinguishing between static and dynamic objects.
In their paper, Max and his team propose a novel hybrid architecture that combines deep learning with radar-based DOGM to improve the differentiation between static and dynamic objects in autonomous vehicles. By introducing a neural network-based state correction mechanism and a heuristic fusion approach, the architecture significantly enhances the detection of slow-moving dynamic objects and overall grid quality, as demonstrated by evaluations using the NuScenes Dataset.
Find here a preprint and all authors of the paper.
3D Object Detection:
ActiveAnno3D Slashes Data Costs and Boosts Efficiency
The curation of large-scale datasets is still costly and requires much time and resources. Data is often manually labeled, and the challenge of creating high-quality datasets remains. In his work, Ahmed and his team fill the research gap using active learning for multi-modal 3D object detection.
At this year’s IEEE IV Symposium, Ahmed presented ActiveAnno3D – the first active learning framework for multi-modal 3D object detection. With this framework, developers can select data samples for labeling that are of maximum informativeness for training.
Summary: Capabilities of ActiveAnno3D
- Performing extensive experiments and ablation studies with BEVFusion and PV-RCNN on the nuScenes and TUM Traffic Intersection datasets and showing that almost the same performance can be achieved when using only half of the data (77.25 mAP compared to 83.50 mAP).
- Integrating active learning framework into the 3D BAT v24.3.2 labeling tool to enable AI-assisted data selection and labeling and minimize the labeling costs.
Find more information on the project page here.
