Beyond the Single Camera: Agentic Multi-View Reasoning in Sports Video Understanding

Noosaga Placements

• Computer Visionsubfield95%
  The paper focuses on multi-view video understanding, a core computer vision task, and introduces a benchmark and method for evaluating MLLMs on sports videos.
  We introduce SportMV-Bench, the first multi-view benchmark designed to evaluate MLLMs on multi-view sports video understanding.We propose SportMV-Agent, an agentic framework that orchestrates an iterative loop of active view selection, perception tool execution, and evidence-grounded reasoning.
• Vision Transformers and Foundation Modelsframework90%
  The paper evaluates and builds upon Vision Transformers and Foundation Models (MLLMs) as baselines and components of the proposed agentic framework.
  Recent Multimodal Large Language Models (MLLMs) achieve strong performance on single-view video understanding benchmarks.We evaluate a broad set of state-of-the-art open-source and proprietary MLLMs under this protocol.
• Active Visionframework85%
  The proposed SportMV-Agent actively selects informative views, which is a form of active vision, to overcome occlusion and gather evidence.
  SportMV-Agent actively decides which view to examine, invokes specialized perception tools to extract concrete evidence, and iteratively accumulates findings until sufficient confidence is reached.Active view selection is critical.
• Deep Learningsubfield80%
  The paper evaluates and builds upon Multimodal Large Language Models (MLLMs), which are deep learning models, and uses them as baselines and components.
  Recent Multimodal Large Language Models (MLLMs) achieve strong performance on single-view video understanding benchmarks.We evaluate a broad set of state-of-the-art open-source and proprietary MLLMs under this protocol.
• Vision-Language Modelsframework70%
  The paper uses Vision-Language Models (MLLMs) as the core technology for video understanding and QA, and the proposed method builds upon them.
  Recent Multimodal Large Language Models (MLLMs) achieve strong performance on single-view video understanding benchmarks.We propose SportMV-Agent, an agentic framework that orchestrates an iterative loop of active view selection, perception tool execution, and evidence-grounded reasoning.
• Natural Language Processingsubfield60%
  The paper uses LLMs for QA generation and verification, and the task involves natural language questions and answers, but the core focus is on video understanding.
  We feed it to an LLM and instruct it to simulate a referee certification exam setter.We employ a strong MLLM (GPT-5) as an auditor.
• Multiagent Systemssubfield50%
  The proposed SportMV-Agent uses an agentic framework with an orchestrator and tools, which aligns with multi-agent systems concepts, but the paper does not focus on multi-agent coordination or communication.
  We propose SportMV-Agent, an agentic framework that orchestrates an iterative loop of active view selection, perception tool execution, and evidence-grounded reasoning.

Abstract

Recent Multimodal Large Language Models (MLLMs) achieve strong performance on single-view video understanding benchmarks. However, sports videos involve dense occlusion, rapid motion, and complex interactions that are difficult to resolve from a single viewpoint. In practice, sports events are recorded from multiple camera angles, providing complementary evidence used by referees. Yet, no existing benchmark evaluates MLLMs on multi-view sports video understanding. To address this gap, we introduce SportMV-Bench, a comprehensive benchmark built from official match recordings, through a dedicated pipeline combining LLM-based generation, MLLM-based verification, and human filtering to ensure quality and consistency. SportMV-Bench containing 787 multi-view video bundles and 2592 question-answer pairs across three categories: Perception-Aware Recognition (PAR), Rule-aware Event Interpretation (REI), and Adjudicative Decision Reasoning(ADR). Our analysis shows that current MLLMs fail to effectively exploit multi-view information, with the bottlenecks lying in fine-grained visual perception and view selection rather than logical reasoning or domain knowledge. We propose SportMV-Agent, an agentic framework that orchestrates an iterative loop of active view selection, perception tool execution, and evidence-grounded reasoning, achieving a significant 14.46% relative improvement over the strongest MLLM baseline.

Paper Context