Files
aimsight-libredesk/snippets/textes/what-should-i-do-if-all-targets-may-potentially-move-or-settle.md
T
maximeandClaude Opus 5.5 10ed887417 Noms plus clairs : snippets/, scripts de test numérotés, verifier_ollama.py
- base_connaissances/ → snippets/ (corpus/ → textes/, instructions.md → consignes_redaction.md, decoupage.py → fichiers.py, importer.py → envoyer.py) ; python3 -m snippets.envoyer.
- instance-de-test/ : 1_demarrer.sh, 2_creer_boite_et_cle.py, 3_brancher_ia.py, 4_essayer_brouillons.py, voir_empreintes.sh, connexion_admin.py.
- ollama/verifier.py → ollama/verifier_ollama.py ; tests, tâches VS Code et docs alignés.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-12 10:27:06 +01:00

1.3 KiB

What should I do if all targets may potentially move or settle?

If no target can be considered perfectly stable, several targets can be configured as references. In some applications, all targets, or preferably a suitable set of relatively distant targets, can be used. Camera motion is estimated using a robust fit over the complete reference network. The algorithm therefore tolerates the displacement of a limited number of individual reference targets without immediately interpreting their displacement as camera motion. The system consequently remains sensitive to differential settlement or deformation within the monitored area. For a real structural displacement to be entirely absorbed into the camera-motion correction, the complete reference network would have to deform according to a very specific geometry that is indistinguishable from a camera rotation. For example, in a simple configuration, all targets would have to settle simultaneously by approximately an amount proportional to their distance from the camera: a target twice as far from the camera would have to settle approximately twice as much. Such a displacement pattern is generally sufficiently unlikely that using a distributed set of potentially moving targets as references remains a robust solution in practice.

Source : AIMsight FAQ, rév. 1.5