AutoMagicCalib Video Calibration
Turn local multi-camera MP4 recordings into AutoMagicCalib calibration results through a REST API.
The skill discloses that videos are uploaded to the AMC backend and requires use on a trusted network; calibration has explicit plan confirmation and actions are scoped to a project. Deductions apply because HTTP is the default, authentication and credential handling are unspecified, rollback/deletion is absent, and configuration files are submitted verbatim.
The documentation and script cover project creation, uploads, verification, polling, and log surfacing, making the happy path coherent. Deductions apply because there is no service-reachability preflight, some GET requests lack raise_for_status, configuration parsing errors are silently ignored, VGGT timeout handling is incomplete, and the script does not fully match the documented confirmation and detector-question behavior.
Trigger phrases, MP4 inputs, naming constraints, optional files, and non-fit cases for RTSP and sample data are clearly described. Deductions apply because Chinese-language interaction and mainland-China environment validation are not provided, and AMC deployment and input-content boundaries remain limited.
The documentation is well layered and includes prerequisites, workflow, outputs, troubleshooting, related skills, version, ownership, and license information; the repository also supplies benchmark and update context. Deductions apply because there is no standalone Examples section, the skill-specific changelog and maintenance/update path are unclear, and the skill-card license wording conflicts with the stated NOASSERTION metadata.
The skill supplies a runnable Python script and an end-to-end REST workflow covering manual alignment, ground truth, focal lengths, and optional VGGT, with project and result locations. Deductions apply because static review cannot verify actual results, the script has preflight and interaction gaps relative to its claims, and the internal benchmark provides only limited evidence, so the static cap prevents a high score.
The source includes an implementation script, evaluation tasks, expected behaviors, and internal benchmark results, providing limited auditability. Deductions apply because no visible CI or committed test suite covers key paths, the benchmark is a single internal evaluation, and independent third-party corroboration is absent.
- The default BASE_URL uses HTTP, and videos or configuration may contain sensitive data; deploy behind a controlled network with authentication and appropriate TLS.
- The script does not perform the documented /v1/ready preflight and starts calibration directly in non-interactive mode; verify reachability and review the plan first.
- Some network responses and VGGT timeouts are not handled completely, so long jobs may leave ambiguous state or created projects.
- There is no evidence of Chinese-language interaction or mainland-China reachability; validate the local AMC service and dependencies before use.
What does this skill do, and when should you use it?
This skill is for users with local prerecorded MP4 files who need to calibrate a new dataset with AutoMagicCalib. It creates a project, uploads videos in alphabetical order, and handles calibration settings, alignment, layout, and optional ground-truth and focal-length files. It verifies the project, starts calibration, polls its state, and retrieves logs and evaluation results. It requires the AMC microservice and UI to be running and does not cover RTSP or other live streams.
Reads local cam_*.mp4 files plus optional settings, alignment, layout, GT.zip, and focal-length inputs; calls the AutoMagicCalib REST API to create a project, upload files, verify readiness, start calibration, and query status; directs the user to the UI when required local files are unavailable; and returns project status, logs, evaluation statistics, and the project ID. If VGGT is ready and the user opts in, it can also run VGGT refinement.
- A computer-vision engineer has synchronized cam_00.mp4, cam_01.mp4, and related recordings and needs a new multi-camera calibration.
- A team has deployed the AutoMagicCalib microservice and UI and wants to submit local videos through the REST API.
- An evaluator has GT.zip and wants average L2 distance and reprojection-error metrics after calibration.
- A user needs to choose the transformer detector for heavier occlusion or the faster default resnet detector.
- A downstream multi-view 3D tracking workflow needs the completed project_id before fetching MV3DT-format results.
What are this skill's strengths and limitations?
- Covers the full REST workflow from project creation through result retrieval.
- Automatically scans for settings, alignment, and layout files while supporting UI fallback.
- Supports resnet and transformer detectors, ground-truth evaluation, focal-length overrides, and optional VGGT refinement.
- Documents important operational decisions such as upload ordering and project-state gates.
- Requires an already running AutoMagicCalib microservice and UI.
- Videos are transmitted to the AutoMagicCalib backend, so deployment must be trusted.
- Supports prerecorded MP4 files only, not RTSP or live streams.
- Calibration typically takes 10–60 minutes, and large uploads may require a longer timeout.
- The source provides no independent execution results or broader platform test matrix for this skill.
How do you install this skill?
Install the skill from NVIDIA's collection with the skills CLI:
npx skills add nvidia/skills --skill amc-run-video-calibration --yes
You can also target a client explicitly:
npx skills add nvidia/skills --skill amc-run-video-calibration --agent codex
The source does not document additional manual installation steps; the collection README says cloning the repository or copying skill folders by hand is unnecessary.
How do you use this skill?
Have the AMC microservice and UI running, then prepare the microservice URL, a project name, and time-synchronized local videos named cam_00.mp4, cam_01.mp4, and so on. Prompt the agent with “calibrate my videos” or “run AMC on these videos” and provide the video directory, microservice URL, and project name. Supply or confirm settings, alignment, layout, detector choice, and optional ground truth or focal lengths as requested. The agent asks for confirmation before calling /calibrate, then polls until COMPLETED, ERROR, or another state requiring action. Use a different skill for RTSP or live-stream requests.
How does this skill compare with similar options?
The source explicitly distinguishes this skill from amc-run-rtsp-calibration: this skill handles local prerecorded MP4 files, while live RTSP streams should use the other skill. It also supports completing missing local-file steps through the AutoMagicCalib UI.