Uncertainty-based Geometry Modeling for Monocular 3D Object Detection
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Lu, YanAbstract
Monocular 3D object detection is a key technology in image-based autonomous driving perception. Due to its challenging of the lack of depth. Existing methods solve it by introducing geometry prior such as estimating depth by perspective projection.
This thesis explores the ...
See moreMonocular 3D object detection is a key technology in image-based autonomous driving perception. Due to its challenging of the lack of depth. Existing methods solve it by introducing geometry prior such as estimating depth by perspective projection. This thesis explores the limitations of the previous geometry-based approaches, providing an error amplification problem suffered by such methods and concludes its specific influence on the error amplification problem as two negative effects: training stability and inference reliability. For the inference reliability problem, a Geometry Uncertainty Projection (GUP) module is proposed to model the perspective projection under the probabilistic modeling framework. It combines the uncertainty modeling and the geometry prior, which can provide a reliable uncertainty to reflect the quality for each projected depth and 3D object detection, making the monocular 3D detector more reliable. To solve the training instability, this thesis further proposes a heuristic Hierarchical Task Learning (HTL) scheme. It treats the learning of perspective projection as a cascaded multi-task learning problem, making the projection method learn more stably. With this, a GUPNet is proposed to combine that GUP module with the HTL scheme to increase the model inference reliability and stabilize the training process simultaneously. Finally, this thesis provides a Geometry Uncertainty Propagation Network (GUPNet++) to further utilize uncertainty in tackling the error amplification. The GUPNet++ introduces the uncertainty propagation theory and provides a new depth uncertainty. With the novel uncertainty, an uncertainty-based optimization and a novel uncertainty-based score scheme are proposed to make the model more reliable and stable than the GUPNet. The findings and extensive experiments in this thesis confirm the effectiveness of the proposed methods in tackling the error amplification problem of the monocular 3D object detection methods.
See less
See moreMonocular 3D object detection is a key technology in image-based autonomous driving perception. Due to its challenging of the lack of depth. Existing methods solve it by introducing geometry prior such as estimating depth by perspective projection. This thesis explores the limitations of the previous geometry-based approaches, providing an error amplification problem suffered by such methods and concludes its specific influence on the error amplification problem as two negative effects: training stability and inference reliability. For the inference reliability problem, a Geometry Uncertainty Projection (GUP) module is proposed to model the perspective projection under the probabilistic modeling framework. It combines the uncertainty modeling and the geometry prior, which can provide a reliable uncertainty to reflect the quality for each projected depth and 3D object detection, making the monocular 3D detector more reliable. To solve the training instability, this thesis further proposes a heuristic Hierarchical Task Learning (HTL) scheme. It treats the learning of perspective projection as a cascaded multi-task learning problem, making the projection method learn more stably. With this, a GUPNet is proposed to combine that GUP module with the HTL scheme to increase the model inference reliability and stabilize the training process simultaneously. Finally, this thesis provides a Geometry Uncertainty Propagation Network (GUPNet++) to further utilize uncertainty in tackling the error amplification. The GUPNet++ introduces the uncertainty propagation theory and provides a new depth uncertainty. With the novel uncertainty, an uncertainty-based optimization and a novel uncertainty-based score scheme are proposed to make the model more reliable and stable than the GUPNet. The findings and extensive experiments in this thesis confirm the effectiveness of the proposed methods in tackling the error amplification problem of the monocular 3D object detection methods.
See less
Date
2024Rights 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 Electrical and Information EngineeringAwarding institution
The University of SydneyShare