Debiased Deep Evidential Regression for Video Temporal Grounding

Kaijing Ma, Haojian Huang, Jin Chen, Haodong Chen, Xianghao Zang, Han Fang, Chao Ban, Hao Sun, Mulin Chen, Xuelong Li

Keywords: Video Temporal Grounding, Uncertainty Quantification, Multi-Modal Fusion, Deep evidential regression, Evidential deep learning

TL;DR: This paper presents SRAM, a novel approach that leverages Evidential Deep Learning to enhance model's robustness and interpretability in Video Temporal Grounding tasks.

Abstract: Existing Video Temporal Grounding (VTG) models perform well in accuracy but often fail to address open-world challenges posed by open-vocabulary queries and out-of-distribution (OOD) videos, which can lead to unreliable predictions. To address uncertainty, particularly with OOD data, we build a VTG baseline using Deep Evidential Regression (DER), which excels in capturing both aleatoric and epistemic uncertainty. Despite promising results, our baseline faces two key biases in multimodal tasks: (1) Modality imbalance, where uncertainty estimation is more sensitive to the visual modality than the text modality; (2) Counterintuitive uncertainty, resulting from excessive evidence suppression in regularization and uneven sample error distribution in conventional DER. To address these, we propose an RFF block for progressive modality alignment and a query reconstruction task to enhance sensitivity to text queries. Additionally, we introduce a Geom-regularizer to debias and calibrate uncertainty estimation. This marks the first extension of DER in VTG tasks. Extensive experiments demonstrate the effectiveness and robustness of our approach. Our code will be released soon.

Primary Area: applications to computer vision, audio, language, and other modalities

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.

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: 6369