vss-deploy-detection-tracking-2d

This NVIDIA skill deploys, operates, debugs, tears down, and drives the REST API of the RTVI-CV 2D detection and tracking microservice, part of the Video Search and Summarization (VSS) blueprint built on DeepStream and Metropolis. It solves the operational problem of bringing up a real-time perception container (warehouse or smart-city use cases, with backends such as GDINO, RT-DETR, and Sparse4D), configuring its pipeline, managing streams, and checking its health and metrics. It unifies two action surfaces in one skill: deploy/operate/debug/teardown of the container, and calling its REST API on a running instance.

The skill uses action routing to pick a single flow per invocation and delegates detail to references and helper scripts (invoked via run_script). A representative configuration workflow copies helper scripts into the running container and applies pipeline settings in one chained call, using docker exec, docker cp, and a scoped rm -rf of a temporary /tmp/scripts directory inside the container to make the copy deterministic; this cleanup is confined to a container-local temp path, not the host. It documents structured output markers to parse, engine-cache handling, batch-size and output-sink updates, and a precise required format for rendering configuration exit boxes. Prerequisites include an active VSS deployment, NGC credentials in NGC_CLI_API_KEY and NVIDIA_API_KEY for image pulls, and curl/jq/Docker. It is Apache-2.0 licensed with published evaluation results.

It targets DevOps and perception engineers running NVIDIA video-analytics pipelines. Use cases include deploying a warehouse or smart-city perception app, adding or removing streams, checking rtvi-cv health, updating batch size or output sink, and tearing the container down. It performs standard Docker/NGC deployment operations with official tooling, not offensive or destructive host-level actions.