Purpose <p>Precise registration of virtual anatomy to the patient is essential for augmented reality (AR) in craniofacial surgery. Traditional marker-based methods lack adaptability in dynamic environments, while soft tissue landmarks are unreliable. We propose a markerless framework using the upper teeth as rigid landmarks for 6D skull pose.</p> Methods <p>The pipeline employs a fine-tuned Segment Anything Model 2 (SAM 2) to segment teeth from monocular images. A pose estimation model is then trained to predict 6D pose directly from these binary masks. Crucially, we utilize a patient-specific training strategy that enables rapid adaptation to new subjects by fine-tuning on synthetic masks generated solely from the patient’s pre-operative intraoral scan.</p> Results <p>Tested on a new dataset comprising 159 images from eight healthy subjects, the proposed method demonstrates high performance across multiple 6D pose estimation metrics, validating the effectiveness of the framework.</p> Conclusion <p>By leveraging patient-specific synthetic data, our approach eliminates the need for large-scale real-world annotations and prevents overfitting, offering a robust, non-invasive solution for surgical navigation.</p>