Unified Dynamic Face Landmark Detection via FPALP

A new research paper on arXiv (2608.10346) proposes a unified approach to face landmark detection (FLD), addressing two major functional limitations: the need for separate network parameters for each N-point benchmark dataset, and the limitation that models trained on N-point datasets only output those N landmarks. The authors introduce Face Part-Anchored Landmark Positions (FPALPs), where each landmark is represented as a progression value from zero to one along a face part's contour. This representation allows all N-point datasets to be unified into a single dataset. The method uses FPALP-based queries, refined with a cross-modality decoder, to predict landmark coordinates. The approach, called Unified Dynamic Face Landmark Detection, aims to enable models to output any number of landmarks dynamically. The paper is available on arXiv and represents a technical advancement in computer vision, with potential applications in augmented reality, animation, and facial recognition. The research is part of ongoing efforts to improve the flexibility and generalizability of FLD models.