Hello Developers!
We recently released VSS 3.1.0 Early Access which contains the following updates:
-
Deliver Real-Time Video Intelligence – accelerated vision AI microservices - fast, feature extraction and visual understanding:
-
Feature extraction with embedding models
-
Image: C-RADIO, SigLIP2
-
Video: Cosmos Embed
-
-
Object detection & tracking models: RT-DETR, GDINO, Sparse4D
-
VLM-powered alerts & dense captions: Cosmos Reason 2, Qwen 3.5, other VLM NIMs, OpenAI APIs Compatibilityle
-
-
Agentic Search (Alpha feature)
-
Search for actions and temporal activities based on video embeddings, as well as object attributes through image embeddings.
-
Breakdown queries, search across multiple embedding spaces, answer fusion, and VLM-based reflection.
-
-
Faster time to solution with an easy-to-customize modular blueprint.
-
Gives developers maximum flexibility and scalability to customize VSS for their use case. Includes a base configuration and 4 add-on agent workflows.
-
Base configuration deploys in less than 5 minutes - 4x faster than previous versions.
-
Report generation workflow: Automatically creates reports based on customizable templates.
-
Long video summarization workflow: Generates concise overview of hour-long videos with time-stamped events
-
Real-time alerts workflow: Uses a VLM to generate alerts that provide the needed details for action
-
Alert verification workflow: Uses a VLM to review and verify CV alerts to eliminate false positives
-
Search workflow: Search stored and streamed video using natural language prompts for actions, events, and attributes.
-
-
Two scalable examples for smart cities and warehouse safety.
- Optimized reference workflows with example data, use case specific prompts and report templates
-
Expanded platform support
- DGX Spark and AGX/IGX Thor now support based VSS configurations for, real-time alerts and verification agents
The latest updates are available on the VSS GitHub Repository and documentation. Happy developing and we would love to hear your feedback!