Overcoming Domain Shift in Lidar-based 3D Object Detection in Urban Contexts
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Tsai, JenAbstract
3D Object Detection is a vital task in 3D scene understanding for many real-world applications such as autonomous driving, robotics and intelligent transport systems. It serves as a fundamental building block for critical downstream tasks such as object tracking and path prediction. ...
See more3D Object Detection is a vital task in 3D scene understanding for many real-world applications such as autonomous driving, robotics and intelligent transport systems. It serves as a fundamental building block for critical downstream tasks such as object tracking and path prediction. However, the deployment of 3D detectors on different lidar types and environments is not a straightforward process. Despite remarkable accuracy on benchmarks, 3D detectors trained on one lidar dataset often suffer a significant drop in performance when tested on a separate lidar or operational environment. This performance degradation can be attributed to a domain shift between the training and testing datasets, stemming from variations in scan pattern across lidars, weather conditions or scenarios. Consequently, 3D detectors tested in unfamiliar testing environments might fail to detect densely observed objects, have misaligned confidence scores, and show an increase in high-confidence false positives. This compromised operational capability renders the detector unfeasible for use. For instance, applying a detector trained on a 32-beam lidar to a 64-beam lidar dataset results in many false positives and missed detections even within 30m range. This is a substantial barrier that hinders the widespread integration of the latest 3D detector research on different types of lidar technologies. This thesis proposes two approaches that effectively leverage existing large-scale labelled datasets for addressing the domain shift: domain-invariant representation and auto-labelling. Through our experiments, we demonstrate that our domain-invariant approach can obtain improved 3D detection performance for different lidar types without requiring re-training. With our auto-labelling framework, we demonstrate that 3D detectors trained with our generated labels achieve state-of-the-art performance, comparable to training with human-annotated labels in Bird’s Eye View (BEV) evaluation.
See less
See more3D Object Detection is a vital task in 3D scene understanding for many real-world applications such as autonomous driving, robotics and intelligent transport systems. It serves as a fundamental building block for critical downstream tasks such as object tracking and path prediction. However, the deployment of 3D detectors on different lidar types and environments is not a straightforward process. Despite remarkable accuracy on benchmarks, 3D detectors trained on one lidar dataset often suffer a significant drop in performance when tested on a separate lidar or operational environment. This performance degradation can be attributed to a domain shift between the training and testing datasets, stemming from variations in scan pattern across lidars, weather conditions or scenarios. Consequently, 3D detectors tested in unfamiliar testing environments might fail to detect densely observed objects, have misaligned confidence scores, and show an increase in high-confidence false positives. This compromised operational capability renders the detector unfeasible for use. For instance, applying a detector trained on a 32-beam lidar to a 64-beam lidar dataset results in many false positives and missed detections even within 30m range. This is a substantial barrier that hinders the widespread integration of the latest 3D detector research on different types of lidar technologies. This thesis proposes two approaches that effectively leverage existing large-scale labelled datasets for addressing the domain shift: domain-invariant representation and auto-labelling. Through our experiments, we demonstrate that our domain-invariant approach can obtain improved 3D detection performance for different lidar types without requiring re-training. With our auto-labelling framework, we demonstrate that 3D detectors trained with our generated labels achieve state-of-the-art performance, comparable to training with human-annotated labels in Bird’s Eye View (BEV) evaluation.
See less
Date
2023Licence
Copyright All Rights ReservedRights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Engineering, School of Aerospace Mechanical and Mechatronic EngineeringAwarding institution
The University of SydneyShare