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

TL;DR

The paper introduces SportMV-Bench, a new benchmark for multi-view sports video understanding using real match recordings and a quality-controlled pipeline. It reveals that current multimodal models struggle with multi-view reasoning due to poor visual perception and view selection, not reasoning ability. An agentic framework, SportMV-Agent, improves performance by 14.46% through iterative view selection and evidence-based reasoning.

Why It Matters

Sports analysis relies on multiple camera angles, but existing AI models fail to leverage this data effectively. This work enables more accurate and referee-like video understanding in dynamic, occluded environments, advancing real-world applications in sports analytics and automated officiating.

Key Findings

• Current multimodal large language models underperform on multi-view sports video tasks due to limitations in fine-grained visual perception and selecting informative camera angles.
• SportMV-Bench, a high-quality benchmark with 787 video bundles and 2592 QA pairs across three reasoning categories, enables systematic evaluation of multi-view understanding.
• SportMV-Agent, an agentic system that actively selects views, runs perception tools, and grounds reasoning in evidence, achieves a 14.46% relative improvement over the best baseline model.