Extending Uncertainty Quantification to New Domains in Computer Vision

Event details
Date | 28.06.2023 |
Hour | 08:00 › 10:00 |
Speaker | Deniz Mercadier |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Antoine Bosselut
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Amir Zamir
Abstract
Deep learning based methods have come to dominate a variety of computer vision tasks with sheer predictive performance. However, their use in safety critical applications remains limited due to the unreliability of their predictions. Uncertainty quantification methods have been developed to allow these models to estimate the reliability of their own predictions. Unfortunately, these methods are usually developed on classification and regression tasks, and cannot be universally extended to more complex ones. One class of such tasks are those related to 3D shape modeling, applicable to a wide variety of problems in medical imaging, manufacturing and simulation. Our research aims to develop tailored approaches for 3D tasks able to quantify uncertainty over whole surfaces in a fine-grained manner, focusing initially on explicit mesh representations.
In this context, we first introduce the basics of uncertainty quantification on classification and regression tasks, along with contemporary challenges based on an extensive survey. Next, we examine a method to estimate uncertainty for monocular depth estimation. We then discuss another recent work on improving the confidence quality of object detection models. These complex examples finally motivate our research goal of developing methods for advanced 3D shape representation tasks, an unexplored direction as of yet, to our knowledge.
Background papers
1. A review of uncertainty quantification in deep learning: Techniques, applications, and challenges
in Information Fusion - December 202. Authors: Moloud Abdar, et al. Link: https://arxiv.org/pdf/2011.06225.pdf
2. Gradient-based Uncertainty for Monocular Depth Estimation in ECCV 2022. Authors: Julia Hornauer, Vasileios Belagiannis
Link: https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136800598.pdf
3. Bridging Precision and Confidence: A Train-Time Loss for Calibrating Object Detection in CVPR 2023 Authors: Muhammad Akhtar Munir,et al.
Link: https://openaccess.thecvf.com/content/CVPR2023/papers/Munir_Bridging_Precision_and_Confidence_A_Train-Time_Loss_for_Calibrating_Object_CVPR_2023_paper.pdf
Exam president: Prof. Antoine Bosselut
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Amir Zamir
Abstract
Deep learning based methods have come to dominate a variety of computer vision tasks with sheer predictive performance. However, their use in safety critical applications remains limited due to the unreliability of their predictions. Uncertainty quantification methods have been developed to allow these models to estimate the reliability of their own predictions. Unfortunately, these methods are usually developed on classification and regression tasks, and cannot be universally extended to more complex ones. One class of such tasks are those related to 3D shape modeling, applicable to a wide variety of problems in medical imaging, manufacturing and simulation. Our research aims to develop tailored approaches for 3D tasks able to quantify uncertainty over whole surfaces in a fine-grained manner, focusing initially on explicit mesh representations.
In this context, we first introduce the basics of uncertainty quantification on classification and regression tasks, along with contemporary challenges based on an extensive survey. Next, we examine a method to estimate uncertainty for monocular depth estimation. We then discuss another recent work on improving the confidence quality of object detection models. These complex examples finally motivate our research goal of developing methods for advanced 3D shape representation tasks, an unexplored direction as of yet, to our knowledge.
Background papers
1. A review of uncertainty quantification in deep learning: Techniques, applications, and challenges
in Information Fusion - December 202. Authors: Moloud Abdar, et al. Link: https://arxiv.org/pdf/2011.06225.pdf
2. Gradient-based Uncertainty for Monocular Depth Estimation in ECCV 2022. Authors: Julia Hornauer, Vasileios Belagiannis
Link: https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136800598.pdf
3. Bridging Precision and Confidence: A Train-Time Loss for Calibrating Object Detection in CVPR 2023 Authors: Muhammad Akhtar Munir,et al.
Link: https://openaccess.thecvf.com/content/CVPR2023/papers/Munir_Bridging_Precision_and_Confidence_A_Train-Time_Loss_for_Calibrating_Object_CVPR_2023_paper.pdf
Practical information
- General public
- Free
Contact
- edic@epfl.ch