vss-generate-video-calibration

Use to run AutoMagicCalib on local MP4s, RTSP, or the bundled sample dataset, and to deploy vss-auto-calibration when needed. Do not use for non-AMC calibration or runtime analytics.

2,759stars321forksUpdated 8/1/2026

Data Science & ML DevOps & CI/CD API Development #deployment #camera-calibration #nvidia #python #computer-vision #rest-api

Security Assessment

Low Risk(78/100)

Detected risks:

External Credential Usage(The workflow requires an NGC API key and performs a docker registry login to pull NVIDIA-hosted images, so a real external credential is read and used during deployment., Handling follows standard practice with a documented reminder not to include secrets in logs or output, so this is credential setup rather than exfiltration.)

Remote Install Script Piped To Shell(As a documented runtime fallback the skill installs the uv tool by downloading its official install script from a remote host and piping it directly into a shell, which executes remotely fetched code., This is a legitimate official installer invoked only when a local Python virtual environment cannot be created, but the pipe-to-shell pattern still runs code fetched at runtime and should be reviewed before use.)

Privileged System Package Install(As a last-resort fallback the documentation suggests installing python3-venv and python3-pip using a privileged apt command, which modifies system packages with elevated permissions.)

Security Score78/100

About vss-generate-video-calibration

This NVIDIA skill runs AutoMagicCalib (AMC) end-to-end for camera calibration within the Video Search and Summarization (VSS) blueprint, and can deploy the AMC microservice when it is not already running. It solves the setup and orchestration problem of multi-camera calibration: it routes a user's request to the correct input mode (local MP4 files, live RTSP streams, or a bundled synthetic sample dataset) and drives the calibration through the microservice REST API, from project creation through verify, calibrate, poll, and results.

The skill is organized around input-routing tables and per-mode reference files, loading only the reference needed for the selected mode plus shared calibration-tail and common-steps snippets. It probes microservice readiness endpoints, creates fresh projects, uploads or references video sources, and reports metrics such as L2 distance and reprojection error for the ground-truth sample dataset. For environment setup it can obtain a large sample zip from a public GitHub LFS repository, extract it into a cache directory, and self-heal a Python runtime by creating a throwaway virtual environment; as documented fallbacks it may install the uv tool via its official remote install script piped to a shell, or install python3-venv via apt with sudo. Deployment of the AMC service and NGC image pulls (docker login to nvcr.io using an NGC API key) are also covered.

It targets computer-vision and MLOps engineers standing up or verifying the VSS AMC calibration pipeline. Stated limitations include needing the matching VSS profile/microservice reachable, NGC rate limits and GPU-memory requirements, and host-dependent concurrency. The main security considerations are operational rather than malicious: a remote install-script piped to a shell as a runtime fallback, and privileged package installation.

FAQ

What input sources does the calibration support?

Three modes: local MP4 files (cam_*.mp4), live RTSP streams, and a bundled synthetic warehouse sample dataset with ground truth for verifying a fresh install.

What credentials are required?

An NGC API key (NGC_CLI_API_KEY) for pulling NGC-hosted models and NIMs; the docs show docker login nvcr.io and re-exporting the key on 401/403 errors. Handle the key as a normal credential and avoid logging it.

Does it install software on my machine?

It can. To self-heal a Python runtime it may create a throwaway virtual environment, and as documented fallbacks it can install the uv tool via its official remote install script piped to a shell or install python3-venv via apt with sudo. Review these steps before running in sensitive environments.

Where does the sample dataset come from?

A roughly 154 MB zip that is not committed to the VSS repo; it is fetched from the public NVIDIA-AI-IOT auto-magic-calib GitHub repository via git-lfs and cached locally rather than committed.

What are the main limitations?

It requires the matching VSS profile/microservice deployed and reachable, NGC pulls may hit rate limits and GPU-memory requirements, and concurrency and storage limits depend on host hardware.