基 于 特 征 解 耦 的 自 监 督 视 线 估 计 方 法

Gaze estimation, as a key technology in eye movement tracking, is widely applied in human-computer interaction. Existing methods generally rely on a large number of facial images annotated with high-precision gaze directions, yet annotating such data incurs substantial human and material costs. This work proposes a self-supervised gaze estimation method based on feature disentanglement, which constructs gaze direction similar image pairs and identity & head pose similar image pairs using pseudo-labels generated by a pre-trained gaze estimation model, and computes their gaze similarity. A gaze disentanglement dual encoder is employed to extract gaze direction-related features and gaze-irrelevant features(such as head pose and identity information)from similar image pairs, respectively. A semantic disentanglement discriminative module then predicts the semantic similarity between different types of features. Meanwhile, cross-entropy reconstruction is used to iteratively update the gaze similarity of these image pairs, thereby achieving the disentanglement of gaze direction-related and gaze-irrelevant features. Experiments demonstrate that the proposed method can effectively disentangle gaze-related features from irrelevant features and achieves superior performance compared to state-of-the-art methods, with a gaze estimation angular error as low as 4.06° on the MPIIFaceGaze dataset.