Purpose of generative self-adversarial learning. (a) Cross-domain general gaze feature.
Framework of generative self-adversarial learning for gaze estimation. (a) Feature encoder. (b) Gaze regression module. GP is global pooling layer. (c) Adversarial reconstruction module. GRL refers to gradient reversal layer. (d) Loss function.In the gaze regression task, gaze regression module performs cooperative optimization with the feature encoder, encouraging to extract more gaze-relevant features. In the image reconstruction task, the generative reconstruction module performs adversarial optimization with the feature extractor, encouraging to extract fewer features from the input image. As a result, the model is guided to learn only the general gaze features, the domain generalization capability is enhanced.
Visualized reconstruction results. The top row shows the original images, the bottom row shows the reconstructed images.
Subjective evaluation results on driver gaze estimation. (a) Results on XJTU-DA dataset. (b) Results on AUC dataset. The orange, blue and purple arrows respectively refer to the estimation of the drivers' gaze directions of GSA-Gaze, CA-Net, and Dilated-Net.