RIDGE: Region-Informed Derivative-Guided Evidence Selection for Long Video Understanding

TL;DR

RIDGE selects frames in long videos using temporal signals, achieving top performance across four benchmarks.

cs.CV 🔴 Advanced 2026-08-31 3 views
Shanqing Xu Meng Luo Mengchen Qian Yuhui Gao Siyue Peng Xiaohan Zhong Xiaojin Zhang Zhongyu Wei Wei Chen Xiang Bai
long video understanding frame selection temporal signal machine learning computer vision

Key Findings

Methodology

RIDGE framework treats the frame-query similarity curve as a temporal signal, using local changes and curvature to partition the timeline into structural regions, applying region-specific selection to preserve event cores, transitions, buildup, aftermath, and contextual frames under a fixed budget. It is a lightweight post-processing step requiring no training or iterative LVLM calls.

Key Results

  • On the Video-MME benchmark, RIDGE improved by 3.1 percentage points on the LLaVA-OV-7B model, reaching 59.6%.
  • On LVBench, RIDGE improved by 12.4 percentage points on the Qwen2.5-VL-7B model, reaching 50.9%.
  • RIDGE achieved the best performance in most settings across four long-video benchmarks and three backbones.

Significance

RIDGE addresses the frame selection problem in long video understanding by identifying and preserving key parts of the temporal signal relevant to the query. This method is significant in both academia and industry as it enhances the efficiency and accuracy of large vision-language models in processing long videos.

Technical Contribution

RIDGE redefines query-aware keyframe selection as reading a query-conditioned temporal signal. By segmenting the similarity curve into structural regions and performing question-aware, region-specific frame selection, RIDGE preserves both event cores and surrounding temporal evidence without requiring training.

Novelty

RIDGE is the first framework to treat the frame-query similarity curve as a temporal signal, offering a new perspective by focusing not only on the magnitude of similarity values but also on their shape, compared to existing methods.

Limitations

  • RIDGE may face computational resource limitations when handling extremely long videos due to the need for similarity computation for each frame.
  • The method relies on precomputed frame-query scores, which may be sensitive to the accuracy of these scores.

Future Work

Future research can explore the applicability of RIDGE to different types of videos and incorporate more contextual information to improve frame selection accuracy. Additionally, the computational efficiency of RIDGE can be further enhanced by optimizing the similarity computation process.

AI Executive Summary

Long video understanding is a challenging field where existing large vision-language models face limitations due to fixed visual-token budgets. The RIDGE framework treats the frame-query similarity curve as a temporal signal, using local changes and curvature to partition the timeline into structural regions, applying region-specific selection to preserve event cores, transitions, buildup, aftermath, and contextual frames under a fixed budget. This method requires no training or iterative LVLM calls and is a lightweight post-processing step.

RIDGE consistently outperforms uniform sampling across four long-video benchmarks and three backbones, especially excelling in scenarios requiring long-horizon and multi-event reasoning. By identifying and preserving key parts of the temporal signal relevant to the query, RIDGE enhances the efficiency and accuracy of large vision-language models in processing long videos.

Despite its strong performance, RIDGE may face computational resource limitations when handling extremely long videos. Future research can explore its applicability to different types of videos and incorporate more contextual information to improve frame selection accuracy. Additionally, the computational efficiency of RIDGE can be further enhanced by optimizing the similarity computation process.

Deep Analysis

Background

Long video understanding is an intersection of computer vision and natural language processing, gaining attention with the development of large-scale vision-language models. Traditional methods often estimate frame-query relevance for frame selection but tend to overlook the temporal structure of similarity sequences. RIDGE offers a new perspective by treating the frame-query similarity curve as a temporal signal.

Core Problem

Long videos contain far more visual content than large vision-language models can process, making frame selection essential. Existing methods often focus only on high-scoring frames, ignoring the temporal structure of the similarity curve, which can lead to missing frames that explain, contextualize, or follow an event.

Innovation

RIDGE's core innovation is treating the frame-query similarity curve as a temporal signal. By using local changes and curvature to partition the timeline into structural regions and applying region-specific selection, RIDGE preserves event cores, transitions, buildup, aftermath, and contextual frames under a fixed budget without requiring training or iterative LVLM calls.

Methodology

  • �� Treat the frame-query similarity curve as a temporal signal.
  • �� Use local changes and curvature to partition the timeline into structural regions.
  • �� Apply region-specific selection to preserve event cores, transitions, buildup, aftermath, and contextual frames.
  • �� Serve as a lightweight post-processing step on precomputed frame-query scores.

Experiments

Evaluated on four long-video benchmarks: Video-MME, MLVU, LongVideoBench, and LVBench. Compared against uniform sampling, score-ranking, or adaptive selectors like AKS, BOLT, and FOCUS, and the diversity-aware selector MDP3. Used three representative LVLM backbones: Qwen2.5-VL-7B, InternVL-3-8B, and LLaVA-OV-7B.

Results

RIDGE achieved the best performance in most settings, especially excelling in scenarios requiring long-horizon and multi-event reasoning. On the Qwen2.5-VL-7B model, RIDGE improved by 12.4 percentage points on LVBench, reaching 50.9%.

Applications

RIDGE can be applied in scenarios requiring long-horizon and multi-event reasoning, such as video question answering, video retrieval, and video summarization. By identifying and preserving key parts of the temporal signal relevant to the query, RIDGE enhances the efficiency and accuracy of large vision-language models in processing long videos.

Limitations & Outlook

RIDGE may face computational resource limitations when handling extremely long videos due to the need for similarity computation for each frame. Additionally, the method relies on precomputed frame-query scores, which may be sensitive to the accuracy of these scores.

Plain Language Accessible to non-experts

Imagine you're watching a long movie but can only choose some scenes to understand the whole story. RIDGE acts like a smart assistant that analyzes each scene's relevance to the story and selects those most important for understanding the plot. It not only focuses on the most obvious climaxes but also chooses scenes that help you understand the background and subsequent developments. This way, you can better grasp the entire story in a limited time.

ELI14 Explained like you're 14

Imagine you're playing a long game but can only choose some levels to play. RIDGE acts like a smart assistant that analyzes each level's relevance to the game's goal and selects those most important for completing the game. It not only focuses on the most obvious climaxes but also chooses levels that help you understand the game's background and subsequent developments. This way, you can better complete the entire game in a limited time!

Glossary

Frame Selection

The process of selecting a subset of frames from a long video for better understanding and processing.

In long video understanding, selecting relevant frames to reduce computational burden.

Similarity Curve

A curve representing the relevance between frames and a query.

Used to analyze the temporal relevance of frames to a query.

Event Core

The key moment in a video most relevant to the query.

RIDGE identifies event cores by detecting peaks in the similarity curve.

Region-specific Selection

A method of selecting frames based on the characteristics of different temporal regions.

RIDGE performs region-specific selection based on the temporal structure of the similarity curve.

Large Vision-Language Models

Large models capable of processing both visual and language inputs.

Core technology used in long video understanding.

Open Questions Unanswered questions from this research

  • 1 How to improve the accuracy of RIDGE's frame selection without increasing computational burden?
  • 2 What is the applicability of RIDGE to different types of videos?
  • 3 How to optimize the similarity computation process to enhance RIDGE's computational efficiency?

Applications

Immediate Applications

Video Question Answering

By selecting key frames relevant to the question, RIDGE can improve the accuracy and efficiency of video question answering systems.

Long-term Vision

Automated Video Editing

RIDGE can be used to automatically select key segments in videos, enabling intelligent video editing.

Abstract

Long videos contain far more visual content than Large Vision-Language Models (LVLMs) can process under a fixed visual-token budget, making frame selection essential. Existing query-aware selectors usually estimate frame-query relevance and build a compact subset from high-scoring frames. Although their mechanisms differ, the similarity sequence is still often treated primarily as values to rank or sample from, rather than as an ordered signal whose shape reflects how query-relevant evidence emerges, peaks, and fades over time. This can obscure frames that explain, contextualize, or follow an event, because such evidence may lie on the rising or falling sides of a nearby relevance peak and receive lower absolute scores. We propose RIDGE, a frame selection framework that reads the frame-query similarity curve as a temporal signal. By using local changes and curvature, RIDGE partitions the timeline into structural regions and applies region-specific selection to preserve event cores, transitions, buildup, aftermath, and contextual frames under a fixed budget. It is a lightweight post-processing step on precomputed frame-query scores and requires neither training nor iterative LVLM calls. Across four long-video benchmarks and three backbones, RIDGE achieves the best performance in most settings and remains competitive in the others.

cs.CV