amc-run-rtsp-calibration

Skill: Calibrate from RTSP Streams

When to Use This Skill

Activate this skill when the user wants to calibrate from live RTSP camera streams. Typical prompts:

• "calibrate RTSP streams" / "calibrate from live cameras"
• "run AMC on RTSP"
• The user provides one or more rtsp://... URLs

VIOS records fixed-duration clips from each stream, the AMC microservice ingests those clips into a project, then the workflow follows the same verification, calibration, polling, and results path as pre-recorded MP4 calibration.

Do not use this skill for local MP4 files already on disk; route those requests to skills/amc-run-video-calibration/SKILL.md. Do not use it for the bundled sample dataset; route that to skills/amc-run-sample-calibration/SKILL.md.

Never reuse files from the bundled sample dataset, extracted sample zip, assets/, or previous projects for RTSP calibration unless the user explicitly provides those paths for this RTSP scene. Similar camera names, stream counts, or cam_00/cam_01 ordering are not evidence that sample alignment, layout, GT, or detector settings apply.

Prerequisites

• AMC microservice and UI running (follow skills/amc-setup-calibration-stack/SKILL.md if needed).
• VIOS is running and reachable from the AMC microservice.
• VIOS_BASE_URL is configured in the AMC microservice environment before capture starts.
• RTSP URLs are reachable from the VIOS host.
• Camera streams have enough moving people/objects for calibration; record at least 2-3 minutes when possible.
• Python 3 with requests installed when using the bundled script.

Data Privacy

RTSP URLs may contain usernames, passwords, hostnames, or network topology. Do not print full RTSP URLs if credentials are embedded. This skill does not handle bearer credentials; if the VIOS deployment requires authentication, stop and hand the user to a manual admin-managed workflow instead of collecting or relaying secrets in chat, scripts, or logs.

What to Ask the User

Required

1. RTSP URLs, one per camera.
2. Camera names, one per stream. Use cam_00, cam_01, ... if the user does not provide names.
3. Recording duration in seconds. Minimum is 60; prefer 120-180 or more when the scene has sparse motion.
4. Microservice URL, for example http://<HOST_IP>:8000 or http://<HOST_IP>:8000/v1.
5. Project name.
6. Calibration asset source for this RTSP scene:
   • a local directory to scan, such as /data/my_rtsp_calib/;
   • explicit paths to settings, alignment, layout, and optional GT files; or
   • confirmation that the user will upload/tune settings and alignment in the AMC UI.

If the user does not provide a local asset source, stop and ask whether they want to provide a path or use UI upload. Give the UI link as http://<HOST_IP>:<AUTO_MAGIC_CALIB_UI_PORT>; the default UI port is 5000.

Auto-Detected or Asked

RTSP clips are recorded by VIOS, so there is no local videos directory to anchor file discovery. Only scan a directory the user explicitly provided for this RTSP scene. If the user provides a settings file path, use that file's directory as the scan directory. If the user provides a calibration asset directory, scan only that directory. Otherwise ask this question before planning uploads or calibration:

Do you have a local calibration asset directory or settings file for these RTSP streams, or should you upload/tune settings and alignment in the AMC UI at http://<HOST_IP>:<AUTO_MAGIC_CALIB_UI_PORT>?

Posting the settings file replaces UI Step 3 and may pin detector or detector_type. If it pins resnet or transformer, pass that same detector to /calibrate. If no settings file pins a detector, ask the user which detector to use; do not silently default to resnet.

Optional

6. sensor_id per stream if the cameras are already registered in VIOS. Leave unset for auto-registration.
7. Ground truth zip (GT.zip) for evaluation metrics.
8. Focal lengths, one per camera.
9. Whether to run VGGT refinement after AMC completes, only when the project reports vggt_state == "READY".

Instructions

The bundled script in scripts/run_rtsp_calibration.py implements this sequence end to end. Use the prose below for decisions, UI fallback, and troubleshooting.

Step 0 - Verify AMC and VIOS

Confirm the AMC microservice is reachable:

curl -sf http://<HOST_IP>:<MS_PORT>/v1/ready

Confirm VIOS is reachable before starting capture. Probe in this order and stop at the first working URL:

If VIOS is not reachable, ask the user to deploy VIOS and provide the base URL. Do not start RTSP capture until ${VIOS_BASE_URL}/vst/api/v1/sensor/list returns 200.

If VIOS is reachable but the AMC microservice is missing VIOS_BASE_URL, do not edit checked-in compose files. Export the variable and relaunch the microservice with a temporary compose override:

A host-shell export alone is not enough after the container is already running; the microservice process must be restarted with VIOS_BASE_URL in its environment.

Step 1 - Create Project

POST /v1/create_project with form field project_name. Save the returned project_id.

Step 2 - Start RTSP Capture

The response can nest session fields under session:

Save session.session_id.

Step 3 - Poll Capture, Then Ingest

Poll every 10 seconds:

Session lifecycle:

When capture reaches COMPLETED, ingest the recorded clips into the AMC project:

After ingest succeeds, the project has video files attached and the rest of the workflow matches the MP4 upload path.

Need to stop early: POST /v1/rtsp/capture/<project_id>/<session_id>/stop. A partial clip can still be ingested if VIOS produced one.

Other session endpoints:

• GET /v1/rtsp/sessions/<project_id> - list sessions for a project.

Step 4 - Upload Settings, Alignment, Layout, and Optional Files

Resolve local files using the anchor-file pattern above. Upload resolved files:

Use only files from explicit user-provided paths or a user-provided calibration asset directory. Do not extract or scan sample data to find fallback settings, alignment, layout, or GT.

If settings are missing, direct the user to UI Step 3: Parameters at http://<HOST_IP>:<AUTO_MAGIC_CALIB_UI_PORT>, then ask which detector to use (resnet or transformer) before calibration. If alignment or layout is missing, direct the user to UI Step 4: Alignment for this project. For RTSP projects, videos are already ingested; do not re-upload videos in the UI fallback.

Before continuing after UI Step 4, verify:

If the AMC stack stores project outputs outside the default projects/ directory, set PROJECTS_DIR explicitly before running the check above. Do not read compose/.env for project paths during this workflow.

Step 5 - Verify, Calibrate, Poll, and Fetch Results

Verify:

The project must return project_state == "READY".

Confirm the plan before calibrating. Summarize:

• Stream count and recording duration.
• Detector: resnet or transformer.
• Settings source: explicit uploaded settings file, user-provided asset directory, or UI Step 3.
• Alignment/layout source: explicit uploaded files, user-provided asset directory, or UI manual adjustment.
• Optional GT and focal-length overrides.

Start calibration:

Poll:

Stop on COMPLETED or ERROR. On error, fetch GET /v1/amc/calibrate/<project_id>/log.

Fetch results:

Only expect evaluation statistics when GT was uploaded.

Step 6 - Optional VGGT Refinement

After AMC calibration completes, read project_info.vggt_state from GET /v1/get_project_info/<project_id>.

• If vggt_state == "READY", ask whether to run VGGT refinement.
• If confirmed, call POST /v1/vggt/calibrate/<project_id>, poll vggt_state, then fetch GET /v1/vggt_results/<project_id>/evaluation_statistics.
• If VGGT is not ready, skip it and explain that AMC calibration is complete.

Complete Python Script

Use the bundled script from the amc-run-rtsp-calibration skill package, not from the auto-magic-calib repo root. If the user points the agent at this skill folder directly instead of installing it, set AMC_RTSP_SKILL_DIR to the directory containing this SKILL.md, or run the command from that directory.

Common environment variables:

Alternative stream input:

Optional env vars are CALIB_ASSET_DIR, CONFIG_FILE, ALIGNMENT_JSON, LAYOUT_PNG, GT_ZIP, FOCAL_LENGTHS, DETECTOR_TYPE, AMC_UI_URL, SSL_VERIFY, RUN_VGGT, REPO_ROOT, and PROJECTS_DIR. SSL_VERIFY defaults to true; only set SSL_VERIFY=false for loopback testing. If the VIOS deployment requires bearer authentication, stop and use a manual admin-managed capture flow instead of routing credentials through this skill.

Success Criteria

• VIOS health probe returns 200.
• Capture session reaches COMPLETED.
• Ingest returns success and project info shows the expected video files.
• verify_project returns READY.
• AMC calibration reaches project_state == "COMPLETED".
• If GT was uploaded, evaluation statistics are returned.
• No RTSP credentials, bearer tokens, NGC keys, or HuggingFace tokens are printed or persisted by the agent.

Key Output Files

Results persist on the AMC server under:

Troubleshooting

Related Skills

• skills/amc-setup-calibration-stack/SKILL.md - start AMC microservice and UI first.
• skills/amc-run-video-calibration/SKILL.md - calibrate from local pre-recorded MP4 files.
• skills/amc-run-sample-calibration/SKILL.md - verify the stack with the bundled sample dataset.