E-DETR: Evidential Deep Learning for End-to-End Uncertainty Estimation in Object Detection
Keywords: object detection, uncertainty estimation, interpretability
TL;DR: End-to-end instance-aware uncertainty estimation in DETR for object detection leveraging evidential deep learning
Abstract: Detection transformers (DETR) have emerged as powerful end-to-end learning frameworks for object detection, directly regressing detection parameters as point estimates. However, these networks often lack the ability to express any uncertainty within their estimates. In this work, we replace the regression of point estimates with the direct learning of the posterior distribution in a sampling-free manner by leveraging deep evidential learning, complementing the end-to-end DETR architecture. We present an instance-aware uncertainty framework by extending evidential deep learning with an IoU-aware loss, jointly modelling both classification and localization uncertainties. Furthermore, we enable the model to leverage its uncertainty for self-calibration, aligning the predicted probabilities with the true likelihood of outcomes, and effectively apply evidential deep learning for the task of imbalanced dense object detection. Our approach is easily extensible and requires only fine-tuning, thus leveraging the pre-training of transformers on large datasets. We conduct extensive experiments on two in-domain and three out-of-domain datasets, demonstrating impressive improvements in generalization performance, especially when fine-tuning on heavily imbalanced datasets characterized by data scarcity.
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1938
Loading