A Fine-Grained Long Video Understanding Method Based on Multimodal Large Language Models

With the continuous accumulation of video data, efficient and accurate video understanding methods have drawn increasing attention. Existing approaches predominantly rely on supervised learning with deep neural networks. However, their performance is often constrained by high annotation costs and limited generalization, making it difficult to handle the complex temporal dependencies and cross-event reasoning required in long video understanding. Multimodal large language models (MLLMs), empowered by vision-language alignment mechanisms, have demonstrated remarkable capabilities in tasks such as image captioning and visual question answering, opening up new possibilities for long video comprehension. Nonetheless, current MLLM-based methods still suffer from imprecise temporal localization and high computational overhead when applied to long videos. To address these challenges, this paper proposes a novel three-stage long video understanding framework that requires no additional training and leverages existing MLLMs in conjunction with external tools. The framework consists of scene segmentation and effective frame extraction, frame-level semantic description generation, and video structural analysis via MLLMs. Specifically, we first extract semantically representative frames by combining shot segmentation and static-dynamic analysis. Under a zero-shot setting, natural language descriptions are then generated for each frame. These descriptions are temporally aligned with their corresponding timestamps to construct structured prompts, which are subsequently fed into an MLLM to enable fine-grained querying. Experimental results show that the proposed method achieves approximately a 7 % improvement over baselines on fine-grained temporal localization tasks, demonstrating strong practical applicability.