VSS Video Auto-Calibration
Run end-to-end AutoMagicCalib calibration for local videos, RTSP cameras, or the bundled sample dataset.
The documentation includes non-echoed credential handling, NGC/Hugging Face token guidance, pre-deployment image checks, pre-calibration confirmation, scoped ACL advice, and service shutdown instructions. Points are deducted because deployment performs Docker pulls and startup and may change host permissions without a consistently explicit pre-action confirmation or complete rollback procedure; the RTSP example sets ssl_verify=false and does not fully explain RTSP credential data flow or sensitive-data handling.
Routing, prerequisite probes, polling, timeouts, error logs, and troubleshooting are comparatively detailed, making the happy path plausible from static inspection. Points are deducted because the workflow depends on many external services, environment variables, containers, and referenced files not independently available here; abnormal HTTP responses, configuration parsing, and environment-specific failures remain incompletely diagnosed. Static calibration caps this dimension at 10.
Input modes, trigger phrases, required files, non-fit boundaries, and deployment-versus-calibration routing are clearly stated. Points are deducted because Chinese-language support is not documented and core deployment depends on NVIDIA NGC, Docker images, and reachable host networking; sample acquisition also depends on GitHub or other overseas services without a mainland-China fallback.
The skill provides purpose, routing, progressive reference material, examples, limitations, troubleshooting, a version, an Apache-2.0 license, and NVIDIA-related repository pointers. Points are deducted because the author field is missing, maintenance ownership and a changelog/update path are not explicit, and some important assumptions are distributed across referenced files.
The skill covers deployment, MP4, RTSP, and sample calibration and reports project IDs, metrics, logs, and overlay locations, so it has practical workflow value. Points are deducted because success depends on GPU hardware, containers, service availability, permissions, and external images/models; the included benchmark reports only 33% overall effectiveness, and static reading cannot verify directly usable outputs. The static cap therefore limits this score.
The files provide a version, repository pointer, API endpoints, success criteria, evaluation manifests, and a benchmark report, offering some auditability. Points are deducted because there is no independently verifiable CI workflow, committed key-path test suite, or multiple corroborating third-party execution sources; the benchmark is primarily project-provided evidence. Static calibration caps this dimension at 5.
- Deployment pulls external images, starts containers, and may modify host permissions with setfacl; confirm the target host, configuration files, and rollback plan before execution.
- The RTSP example uses ssl_verify=false and may carry username/password credentials in URLs; do not place real credentials in logs, shell history, or shared output.
- Reachability, licensing, and quotas for NGC, GitHub, Hugging Face, and Astral uv require separate validation, particularly on mainland-China networks.
- The benchmark reports 33% effectiveness, and no workflow was executed in this review; manually inspect overlays, logs, and project files before relying on results.
What does this skill do, and when should you use it?
This skill provides the video-calibration workflow for NVIDIA’s Video Search and Summarization Blueprint. It runs AutoMagicCalib against local MP4 files, RTSP streams, or the bundled sample dataset, and can deploy the vss-auto-calibration microservice when required. Through the REST API, it creates and verifies projects, starts calibration, polls progress, retrieves results, and optionally runs VGGT refinement. It requires a matching VSS profile, a reachable service, a writable projects directory, and Python requests.
Routes a request to deployment, local-video, RTSP, or sample-dataset handling and deploys the AMC microservice when needed. It reads video files, RTSP URLs, optional JSON calibration settings, layout and alignment files, and optional GT.zip data; then uses the AMC REST API to create and verify projects, upload or ingest video, start calibration, and poll status every 10 seconds. After completion it retrieves project state, evaluation statistics, PNG overlays, and logs, and can run VGGT refinement when the model is ready. It returns a project_id that downstream consumers can use to download AMC or VGGT MV3DT ZIP output.
- A VSS user needs to calibrate local multi-camera files named cam_00.mp4, cam_01.mp4, and so on.
- A deployment operator needs to calibrate live camera feeds supplied as RTSP URLs.
- A user needs to verify an AMC installation with the bundled sample dataset.
- An operator needs to launch the AMC microservice before calibrating a video collection.
What are this skill's strengths and limitations?
- Supports local MP4, RTSP, and bundled sample-dataset input modes.
- Covers deployment, upload or ingestion, verification, calibration, polling, result review, and troubleshooting.
- Supports resnet and transformer detectors, ground-truth evaluation, VGGT refinement, and MV3DT export.
- Documents common handling for connection failures, NGC authentication errors, GPU-memory failures, and upload timeouts.
- Requires a matching VSS profile, the AMC microservice, a writable projects directory, and network access.
- Calibration commonly takes 10–60 minutes; the bundled sample typically takes 10–30 minutes.
- NGC-hosted models and NIMs may involve rate limits, GPU-memory requirements, and license restrictions.
- The supplied material does not specify a complete supported-host platform list or hardware matrix.
How do you install this skill?
Install the individual skill with the skills CLI:
npx skills add nvidia/skills --skill vss-generate-video-calibration --yes
The repository README states that the CLI runs through npx and prompts for an installation destination; cloning the repository or manually copying the skill directory is not required.
How do you use this skill?
Prompt a compatible agent with requests such as “calibrate my videos” or “calibrate RTSP streams.” To deploy the service, use “launch AMC.” The skill selects the appropriate workflow; combined requests deploy first and then calibrate. After input preparation, it verifies that the project is READY, confirms the detector and calibration settings, calls /v1/calibrate/<project_id>, polls until COMPLETED, and provides the UI, overlay, project-file, and log locations. If settings, layout, or alignment files are missing, follow the instructions to save the required content in the AMC UI.
How does this skill compare with similar options?
The source explicitly names a UI-driven workflow as an alternative: when settings, layout, or alignment files are unavailable, those steps can be completed in the AMC UI before continuing with API calibration.