DeCafNet: Delegate and Conquer for Efficient Temporal Grounding in Long Videos

TL;DR

DeCafNet addresses the computational bottleneck of long video temporal grounding by introducing a lightweight sidekick encoder that densely processes all clips and generates a saliency map to select a small subset for the heavyweight expert encoder. The DeCaf-Grounder then fuses dense and salient features with query information through query-aware temporal aggregation and multi-scale refinement to accurately localize moments. Across Ego4D-NLQ, Ego4D-Goalstep, and MAD, DeCafNet achieves state-of-the-art grounding while reducing computation by up to 47%, demonstrating strong efficiency and performance gains. The approach establishes a practical framework for scalable LVTG in long videos with potential broad impact on retrieval systems and surveillance applications.

Abstract

Long Video Temporal Grounding (LVTG) aims at identifying specific moments within lengthy videos based on user-provided text queries for effective content retrieval. The approach taken by existing methods of dividing video into clips and processing each clip via a full-scale expert encoder is challenging to scale due to prohibitive computational costs of processing a large number of clips in long videos. To address this issue, we introduce DeCafNet, an approach employing ``delegate-and-conquer'' strategy to achieve computation efficiency without sacrificing grounding performance. DeCafNet introduces a sidekick encoder that performs dense feature extraction over all video clips in a resource-efficient manner, while generating a saliency map to identify the most relevant clips for full processing by the expert encoder. To effectively leverage features from sidekick and expert encoders that exist at different temporal resolutions, we introduce DeCaf-Grounder, which unifies and refines them via query-aware temporal aggregation and multi-scale temporal refinement for accurate grounding. Experiments on two LTVG benchmark datasets demonstrate that DeCafNet reduces computation by up to 47\% while still outperforming existing methods, establishing a new state-of-the-art for LTVG in terms of both efficiency and performance. Our code is available at https://github.com/ZijiaLewisLu/CVPR2025-DeCafNet.

Figures (4)

• Figure 1: Model inference time and grounding performance on Ego4d-NLQ ego4dnlq-hou2023groundnlq dataset using one A100 gpu. The circle sizes indicate the TFLOPs for methods. Numbers beside the green circles indicate the amount of salient clips processed by expert encoder.
• Figure 2: DeCafNet Overview. The sidekick encoder efficiently extracts features from input video clips, which, combined with text features, generate a saliency map to select the most salient clips. The expert encoder then processes these salient clips. DeCaf-Grounder uses the extracted features from both the sidekick and the expert encoders to predict the moment associated with the input text query.
• Figure 3: DeCafNet's qualitative results, where Ours wo DCG and Ours w DCG indicate predictions without and with DeCaf-Grounder, respectively. The second row displays the generated saliency maps. DeCafNet yields accurate saliency maps and better grounding results.
• Figure 4: Details of multi-scale temporal refinement. The multi-scale features produced by the temporal transformer are transformed into grounding scores using an FFN classifier. To synchronize grounding information across different scales, we utilize linear interpolation and temporal convolution. Finally, average pooling is applied to effectively combine the synchronized features with the input features.