L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Développement · SKILL.md

video-and-tracking

video-and-tracking

Run 3DDFA_V2 video, smoothing, and manual webcam tracking workflows.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: "video-and-tracking" description: "Run 3DDFA_V2 video, smoothing, and manual webcam tracking workflows." metadata: disco-role: operating disable-model-invocation: true license: MIT

Video and tracking

Use this sub-skill for 3DDFA_V2 video-file processing and smoothing. Webcam support exists in the source repo but remains a manual, hardware-dependent workflow.

When to read

Read this sub-skill when the task asks to:

  • Run demo_video.py or demo_video_smooth.py on an AVI/MP4 file.
  • Produce tracked 2d_sparse, 2d_dense, or 3d video outputs.
  • Tune n_pre, n_next, start, or end frame-window settings.
  • Diagnose imageio, ffmpeg, tracking drift, or webcam/display issues.

Before running

  1. Build and smoke-test the checkout with ../setup-and-assets/.
  2. Confirm imageio and imageio-ffmpeg are installed.
  3. Start with the sample clip if the user's input format is uncertain.

Basic video wrapper

The basic wrapper preserves the original demo_video.py CLI. Put original arguments after --:

python <skill-root>/sub-skills/video-and-tracking/scripts/run-video.py \
  --repo-root <checkout> -- \
  -f <video-path> -o 3d --onnx

Smoothed video wrapper

The smoothed wrapper preserves the original demo_video_smooth.py CLI:

python <skill-root>/sub-skills/video-and-tracking/scripts/run-video-smooth.py \
  --repo-root <checkout> -- \
  -f <video-path> -o 2d_sparse -n_pre 1 -n_next 1 --onnx

Tracking behavior

  • The first frame uses detection; later frames track from the previous landmark state.
  • If the ROI area becomes too small, the demo re-runs face detection.
  • Motion that is too fast or head pose beyond roughly 90 degrees can make the lightweight tracker fail.
  • Smoothing uses a simple average over n_pre + n_next + 1 frames and drops or pads boundary frames.

Webcam boundary

The repo has a webcam script that uses imageio.get_reader("<video0>") and cv2.imshow. This skill documents the workflow in references/workflows.md, but does not provide an automated webcam wrapper because it needs live camera hardware, GUI display access, and manual quit handling.

Troubleshooting

Read references/troubleshooting.md for video-specific failure recovery and ../../references/troubleshooting.md for shared build/import failures.

Skills enregistrées