What does this skill do, and when should you use it?
This skill validates an NVIDIA AutoMagicCalib microservice by running calibration on the shipped sample dataset. The dataset contains four synthetic warehouse camera streams, alignment data, a layout image, and ground truth. The workflow creates a project, uploads the files, starts calibration, polls for completion, and reads evaluation metrics. It requires an already running AMC service and the sample archive at the documented location.
Scans ports 8000–8009 for a successful /v1/ready response; checks for and extracts assets/sdg_08_2_sample_data_010926.zip; creates a project and uploads four name-sorted videos, alignment data, the layout image, and GT.zip; verifies the project, starts resnet calibration, and polls until completion; attempts VGGT refinement when the project reports READY; reads evaluation statistics and identifies the final calibration files on the server. It also documents the equivalent manual Swagger UI sequence.
- An engineer has just installed AMC and wants an end-to-end smoke test using NVIDIA's bundled sample.
- A developer has an AMC microservice running and needs to verify project creation, uploads, and calibration APIs.
- An integration tester wants L2 distance and reprojection-error checks from a four-camera synthetic warehouse dataset with ground truth.
- A user who prefers manual interaction needs a Swagger UI walkthrough instead of the Python script.
How do you install this skill?
- Execution creates an AMC project and uploads sample files to the microservice; confirm the target service, permissions, and retention policy and prepare a manual project-deletion rollback.
- The script may install requests into a temporary virtual environment and recommends sudo apt when venv support is missing, requiring additional system privilege and network access.
- Some metrics, thresholds, and VGGT behavior are stated or reported by internal evaluation rather than independently reproduced here; representative metrics are not guarantees.
- The default is an unauthenticated-looking HTTP localhost API, with no TLS, authentication, or cross-host exposure guidance.
- Shell / CLI
- Network access
- Local filesystem
AMC microservicePython 3requestscurlunzip
Install the NVIDIA skill collection with: npx skills add nvidia/skills. Install this skill directly with: npx skills add nvidia/skills --skill amc-run-sample-calibration --yes. For Codex-specific installation, use: npx skills add nvidia/skills --skill amc-run-sample-calibration --agent codex. The source does not document a complete standalone folder-copy installation procedure.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/amc-run-sample-calibration" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- test the sample dataset
- run sample calibration
After installation and skill loading, prompt the agent with “test the sample dataset” or “run sample calibration”. Prepare a running AMC microservice, an AutoMagicCalib checkout, and assets/sdg_08_2_sample_data_010926.zip; set REPO_ROOT to the checkout, then run the bundled script from the skill directory: python3 scripts/run_sample_calibration.py. If the service is not running, use amc-setup-calibration-stack first. Alternatively, open http://<HOST_IP>:<MS_PORT>/docs and follow the documented order: create a project, upload the four videos, upload alignment/layout/GT files, verify the project, calibrate, poll project status, and read evaluation_statistics.
What are this skill's strengths and limitations?
- Covers project creation, uploads, calibration, evaluation, and optional VGGT refinement.
- The sample includes ground truth, enabling L2-distance and reprojection-error metrics.
- Backend discovery, extraction, and common failure handling are explicitly documented.
- Supports both a Python script and a manual Swagger UI workflow.
- 限定 to the fixed sdg_08_2_sample_data_010926.zip; it is not for user-owned videos or RTSP streams.
- Requires a running AMC microservice and a valid AutoMagicCalib checkout.
- The source provides no actual run record, complete test suite, or cross-platform test evidence.
- Calibration may take about 30 minutes, and timeout diagnosis depends on service logs.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| AMC Sample Dataset Calibration this page ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| AutoMagicCalib Video Calibration ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| FastAPI Router Builder ✓ Microsoft · Official | 51 · Use with care | ★ 2.7k | 3mo ago | MIT |
| Azure Short-Audio Speech-to-Text REST ✓ Microsoft · Official | 49 · Use with care | ★ 2.7k | 3mo ago | MIT |
| cuOpt REST Server Deployment & Python Client ✓ NVIDIA · Official | 45 · Not recommended | ★ 3.8k | today | Apache-2.0 |
Use amc-run-video-calibration for user-owned prerecorded MP4 files, amc-run-rtsp-calibration for live RTSP streams, and amc-setup-calibration-stack when the AMC microservice has not been started.
How did FollowSkills review this skill?
The skill scopes execution to a fixed synthetic sample, explicitly excludes user videos and RTSP, and forbids fabricated results; the script uses a throwaway virtual environment and does not require credentials. Eight points are deducted because it creates projects, uploads data, and persists server-side results without pre-execution confirmation, deletion/rollback guidance, or retention disclosure; HTTP transport, server authorization boundaries, and sensitive-data handling are also under-specified.
The documentation and script are broadly consistent and include sorted uploads, polling, timeouts, terminal-error handling, and log-tail feedback. Twelve points are deducted because static files do not prove the key path runs; the script lacks a backend preflight and clear network-exception handling, dependency installation may fail, and partial uploads have no cleanup path. The static ceiling also prevents a score above 10.
Trigger phrases, the fixed dataset, prerequisites, and routing boundaries to video/RTSP skills are clear, with both Python and Swagger paths provided. Four points are deducted for no Chinese-language guidance or localized failure messages, dependence on a local AMC service and CLI environment, and no mainland-China network or remote-deployment fit discussion; “launch and test” also depends on another skill.
The package has a stable name, version, author, owner, license, metadata, prerequisites, troubleshooting, related skills, and evaluation files, with generally good layering. Five points are deducted for the missing recommended Examples section, absent explicit changelog and stable parameter contract, unclear maintenance/update path, and an empty References section in the skill card. The license metadata also conflicts at governance level with the stated NOASSERTION.
The goal is to validate the AMC workflow end to end with a fixed sample; the API order, GT metrics, and output locations are concrete, and the committed evaluation report provides limited evidence of core value. Nine points are deducted because no independently reproducible actual output or statically verifiable representative result is included, manual UI/service inspection remains necessary, and VGGT and external-service dependencies make the full outcome variable. The static ceiling limits this score to 7.
The SKILL.md, script, evaluation tasks, and benchmark report provide auditable material and distinguish expected outcomes from execution prerequisites. Seven points are deducted because there is no committed test suite or real CI execution evidence covering the key paths, and the benchmark cannot be independently reproduced from the supplied files alone. The static ceiling limits this score to 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →