摘要
Video Temporal Grounding aims to localize specific events within untrimmed videos based on natural language queries. While conventional methods assume the target event is always present, real-world applications necessitate the ability to reject ”ungroundable” queries where the target segment is absent. Current approaches typically rely on Multiple Instance Learning and binary classification, which often struggle to capture nuanced event semantics, leading to unreliable predictions: low-confidence results for groundable queries due to background interference, and paradoxically overconfident predictions for ungroundable ones. To address these challenges, this paper proposes a novel Evidential Learning framework for Uncertainty-Aware Temporal Video Grounding and Rejection (ELVTG). By explicitly quantifying epistemic uncertainty, ELVTG distinguishes groundable queries from ungroundable ones through evidence-driven reasoning. Specifically, we introduce: (1) a foreground evidence collection mechanism that aggregates discriminative semantics via evidential deep learning to mitigate noise; (2) an adaptive evidential learning strategy that dynamically guides the model from contextual understanding to informative evidence gathering; and (3) an uncertainty penalty loss to resolve confidence-uncertainty misalignment. Extensive experiments on the Charades-RF and ActivityNet-RF benchmarks demonstrate that ELVTG achieves state-of-the-art performance in both grounding accuracy and rejection robustness.
| 源语言 | 英语 |
|---|---|
| 期刊论文编号 | 104967 |
| 期刊 | Computer Vision and Image Understanding |
| 卷 | 273 |
| DOI | |
| 出版状态 | 已出版 - 1 11月 2026 |
学术指纹
探究 'Evidential learning for Uncertainty-Aware Video Temporal Grounding and Rejection' 的科研主题。它们共同构成独一无二的学术指纹。引用此
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver