Point Transformer 와 적응형 랜드마크 선별 추론을 이용한 3 차원 얼굴 랜드마크 검출 방법 = 3D Facial Landmark Detection Method Based on Point Transformer and Adaptive Landmark Selection Inference

다국어 초록 (Multilingual Abstract)

Facial landmark detection serves as a fundamental component in a wide range of applications, such as expression analysis, face recognition, and 3D avatar generation. In particular, 3D point cloud–based methods exploit depth information to achieve higher geometric fidelity. Despite this advantage, they still suffer from point misalignment and ambiguous landmark boundaries, hindering precise landmark localization. To address these issues, we propose a 3D facial landmark detection method that integrates Point Transformer V3–based feature extraction with region-adaptive refinement. These probabilities are then used to obtain initial landmark coordinates by performing weighted aggregation. Next, the Region Weight Adapter (RWA) divides the face into eight semantically meaningful facial regions. It then learns region-specific importance weights in conjunction with multi-head self-attention over landmarks. This

design emphasizes informative regions while suppressing cross-region interference. Finally, a GCN- based refinement module takes the initial coordinates, region embeddings, and refined landmark features

as node attributes and predicts residual corrections to enforce structural consistency among neighboring landmarks. We evaluate the landmark detection performance using the AI Hub Korean 3D facial scanning dataset. The proposed model outperformed the conventional PAConv model, achieving a 7.1% improvement in normalized error. This confirms the effectiveness of our approach in increasing the accuracy of 3D facial landmark detection.