FlashVTG: Feature Layering and Adaptive Score Handling Network for Video Temporal Grounding

TL;DR

FlashVTG addresses VTG challenges by replacing decoders with a Temporal Feature Layering ($TFL$) pyramid and an Adaptive Score Refinement ($ASR$) module to capture multi-scale video content and context-aware confidence. The method fuses video and text through an Adaptive Cross Attention mechanism with dummy tokens, processes features with multi-scale layers, and refines predictions via intra- and inter-scale score fusion, augmented by a Clip-Aware Score loss. Experiments across five VTG datasets show state-of-the-art performance in both Moment Retrieval and Highlight Detection, including large gains for short moments, without additional training burden. The approach offers practical impact for precise moment localization in long videos and robust cross-modal ranking, with publicly available code for replication.

Abstract

Text-guided Video Temporal Grounding (VTG) aims to localize relevant segments in untrimmed videos based on textual descriptions, encompassing two subtasks: Moment Retrieval (MR) and Highlight Detection (HD). Although previous typical methods have achieved commendable results, it is still challenging to retrieve short video moments. This is primarily due to the reliance on sparse and limited decoder queries, which significantly constrain the accuracy of predictions. Furthermore, suboptimal outcomes often arise because previous methods rank predictions based on isolated predictions, neglecting the broader video context. To tackle these issues, we introduce FlashVTG, a framework featuring a Temporal Feature Layering (TFL) module and an Adaptive Score Refinement (ASR) module. The TFL module replaces the traditional decoder structure to capture nuanced video content variations across multiple temporal scales, while the ASR module improves prediction ranking by integrating context from adjacent moments and multi-temporal-scale features. Extensive experiments demonstrate that FlashVTG achieves state-of-the-art performance on four widely adopted datasets in both MR and HD. Specifically, on the QVHighlights dataset, it boosts mAP by 5.8% for MR and 3.3% for HD. For short-moment retrieval, FlashVTG increases mAP to 125% of previous SOTA performance. All these improvements are made without adding training burdens, underscoring its effectiveness. Our code is available at https://github.com/Zhuo-Cao/FlashVTG.

Figures (2)

• Figure 1: Comparison of Model Performance on the Moment Retrieval Task using video query pair from QVHighlights lei2021detecting. Ground Truth consists of two short moments, both of which are precisely retrieved by our model. In contrast, Moment-DETR lei2021detecting, the established benchmark, and R2-Tuning liu2024tuning, the previously leading method, failed to accurately retrieve the designated moments.
• Figure 2: Overview of the FlashVTG Framework. As depicted in the blue section below, input videos and queries are first processed through frozen encoders to extract corresponding video and text features. These text features, concatenated with encoded Dummy tokens to form the Key, are merged with video text features in the Feature Fusion module to create Fused Features. These are then directed into the Temporal Feature Module and the HD Head, producing $K$ temporal scale features ($f^{i}_j$ refers to the token at the $j^{th}$ position of the feature $F_i$ at the $i^{th}$ scale.) and saliency scores, respectively. These features and their concatenated forms are input into the Adaptive Score Refinement Module, generating intra- and inter-scale confidence scores $c$ (shown here with three examples). Lastly, the Moment Retrieval Head uses all fused features for boundary prediction, outputting timestamps $(b_s, b_e)$ for the start and end points.