Video model improves summaries for long videos with smarter scene splitting
Long-form Video-to-Text Summarization using Recursive Language Models
Computer Vision and Pattern Recognition
Summary
Videos that last a long time are hard to summarize well because it’s tricky to keep track of the story throughout. The authors created a new video language model that breaks long videos into meaningful parts based on scenes, helping it keep the story clear. It also smartly calls on extra tools like speech and object recognition to add important details to the summary. Tests show this model creates better, more accurate summaries than other methods, especially by naming important people and events correctly.
What this means in practice
- •For video content teams: Generate clear, detailed summaries of long videos by splitting them into scenes and enriching descriptions with recognized actions and speech.
- •For security monitoring teams: Produce accurate textual summaries for long surveillance footage by combining scene-based video sections with automated recognition tools for key events.
Authors
Alessia Intini, Lorenzo Vaiani, Enrico Chen, Luca Cagliero
Abstract
While VideoLLMs have made significant progress in understanding short video benchmarks, they often struggle to summarize long-form ones. This is mainly due to performance degradation caused by long-context prompts, the inherent difficulty of preserving temporal continuity and a lack of semantically rich concepts in event descriptions. To tackle these issues, this paper introduces the first recursive Video Language Summarization Model (RVLSM) designed for video processing and tailored to long-form video-to-text summarization. RVLSM recursively decomposes videos into semantically coherent segments and delegates chunk-level reasoning to VideoLLM sub-models and external tools. It employs (1) Semantic Temporal Decomposition (STD), which preserves temporal coherence and continuity through a scene-aware chunking strategy that dynamically partitions videos according to semantic boundaries, and (2) Adaptive Multimodal Tool Calling (AMTC), which ensures semantic richness of the summary via a routing mechanism that selectively invokes auxiliary modules such as action recognition, optical character recognition, automatic speech recognition, and object detection. Experiments on the VideoXum and MixTyVSum benchmarks show that RVLSM consistently improves over zero-shot VideoLLM baselines across multiple model settings and performance metrics, and is competitive with a state-of-the-art fine-tuned model for video-to-text summarization. Importantly, the named entities in the RVLSM's summaries align more closely with those detected in the ground-truth summaries than those detected by baseline and state-of-the-art methods, confirming the saliency of the event descriptions it generates.