NVIDIA Enhances Video Analytics with Generative AI and Reasoning

NVIDIA has released its Video Search and Summarization (VSS) Blueprint 2.4, leveraging generative AI and reasoning models to revolutionize video analytics. This upgrade enables real-time understanding of video content, moving beyond simple object counting to delivering actionable insights.

Key Enhancements in VSS Blueprint 2.4

1. Improved Physical World Understanding: Integration with NVIDIA Cosmos Reason, a 7-billion-parameter reasoning Vision Language Model (VLM), enhances scene understanding and physical AI reasoning.
2. Enhanced Q&A Capabilities: New knowledge graph features include entity deduplication, agentic graph traversal, and support for Neo4J and ArangoDB with CUDA acceleration.
3. Edge Deployment with Event Reviewer: A lightweight API allows VSS to augment existing computer vision pipelines, providing generative AI insights for key events.
4. Expanded Hardware Support: VSS now runs on NVIDIA Blackwell platforms, including Jetson Thor, RTX Pro 6000, and DGX Spark.

Figure 1. The VSS architecture combines ingestion and retrieval pipelines for efficient video analysis.

Why This Matters

• Accuracy Boost: Benchmark tests show a 16.16% improvement in LongVideoBench accuracy and 10.20% in MLVU.
• Cross-Camera Insights: Knowledge graph deduplication enables tracking objects across multiple cameras.
• Cost Efficiency: The Event Reviewer feature reduces compute costs by focusing generative AI on critical clips.

Figure 2. Deduplication streamlines the knowledge graph for better Q&A accuracy.

Practical Applications

• Manufacturing: Identify bottlenecks or safety violations.
• Retail: Analyze customer behavior across store cameras.
• Transportation: Monitor traffic flow and incident root causes.

Getting Started

Developers can deploy VSS via NVIDIA Brev Launchables or explore the GitHub repo for training notebooks and reference code.

For production deployments, refer to the VSS documentation.