Explicit temporal inductive biases for video moment retrieval via multi-scale Gaussian priors
- 1China Telecom Research Institute, Beijing, China
Abstract
Video moment retrieval aims to localize video segments corresponding to natural language queries. While Transformer models are widely applied, their localization performance declines when processing diverse temporal scales. This limitation arises from the dependence on implicitly learned decoder queries. Without explicit temporal grounding, these unstructured embeddings lead the decoder to perform a global search across the full video sequence. We propose Gaussian-DETR to address this by introducing explicit temporal awareness into the decoding process. Our framework replaces unstructured embeddings with queries constructed from multi-scale Gaussian temporal priors, parameterized by a learnable center and scale. To adapt these priors to the input, we design a multi-scale Gaussian pooling mechanism that extracts local features from the encoder for query initialization, introducing a temporal inductive bias while maintaining the architectural efficiency of DETR. On the QVHighlights benchmark, Gaussian-DETR provides a scale-specific benefit for long-event retrieval, improving Long mAP by +1.98 points over Moment-DETR with statistical support (p = 0.0111), while the differences in overall recall are statistically inconclusive. Ablations suggest that the Gaussian prior improves scale-aware query initialization by balancing long-event coverage and local temporal specificity. This mechanism provides a parameter-efficient component for studying scale-aware query initialization in DETR-like video moment retrieval models.
Related
- Investigating the role of the hippocampus in the implicit and explicit expression of statistical learning in patients with temporal lobe epilepsy
- Dynamic mechanical sorting biases niche occupancy and lineage outcome during development of bilayered glandular epithelia
- Factors influencing tuberculosis preventive treatment implementation in India: A hybrid inductive-abductive thematic analysis using implementation framework
- Mitigating biases of rescaling in forward-in-time population genetic simulations
- Cophylogeny simulators are not interchangeable: similarities, differences and structural biases in synthetic host-symbiont
BibTeX
@article{dai2026explicit,
title = {{Explicit temporal inductive biases for video moment retrieval via multi-scale Gaussian priors}},
author = {Guanlin Dai and Da Gao and Chao Zheng},
journal = {PLOS One},
year = {2026},
month = sep,
doi = {10.1371/journal.pone.0354458},
url = {https://doi.org/10.1371/journal.pone.0354458},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Explicit temporal inductive biases for video moment retrieval via multi-scale Gaussian priors
AU - Guanlin Dai
AU - Da Gao
AU - Chao Zheng
PY - 2026
DA - 2026/09/01
JO - PLOS One
DO - 10.1371/journal.pone.0354458
UR - https://doi.org/10.1371/journal.pone.0354458
AB - Video moment retrieval aims to localize video segments corresponding to natural language queries. While Transformer models are widely applied, their localization performance declines when processing diverse temporal scales. This limitation arises from the dependence on implicitly learned decoder queries. Without explicit temporal grounding, these unstructured embeddings lead the decoder to perform a global search across the full video sequence. We propose Gaussian-DETR to address this by introducing explicit temporal awareness into the decoding process. Our framework replaces unstructured embeddings with queries constructed from multi-scale Gaussian temporal priors, parameterized by a learnable center and scale. To adapt these priors to the input, we design a multi-scale Gaussian pooling mechanism that extracts local features from the encoder for query initialization, introducing a temporal inductive bias while maintaining the architectural efficiency of DETR. On the QVHighlights benchmark, Gaussian-DETR provides a scale-specific benefit for long-event retrieval, improving Long mAP by +1.98 points over Moment-DETR with statistical support (p = 0.0111), while the differences in overall recall are statistically inconclusive. Ablations suggest that the Gaussian prior improves scale-aware query initialization by balancing long-event coverage and local temporal specificity. This mechanism provides a parameter-efficient component for studying scale-aware query initialization in DETR-like video moment retrieval models.
SN - 1932-6203
ER -