Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation

TL;DR

This work identifies a previously overlooked source of error in cross-domain gaze estimation: gaze label distribution shift arising from label acquisition and individual physiology. It introduces Gaze Label Alignment (GLA), a pre-processing pipeline that aligns labels across multiple source domains by training an anchor-domain regressor, predicting labels on other domains, and fitting a mapping to align their labels with the anchor. GLA is designed to be compatible with any gaze estimation method and can boost state-of-the-art DG/DA approaches by reducing label misalignment, as demonstrated across multiple gaze datasets with consistent performance gains. The method yields practical impact by improving cross-domain gaze estimation robustness in real-world, privacy-preserving, and multi-source settings.

Abstract

Gaze estimation methods encounter significant performance deterioration when being evaluated across different domains, because of the domain gap between the testing and training data. Existing methods try to solve this issue by reducing the deviation of data distribution, however, they ignore the existence of label deviation in the data due to the acquisition mechanism of the gaze label and the individual physiological differences. In this paper, we first point out that the influence brought by the label deviation cannot be ignored, and propose a gaze label alignment algorithm (GLA) to eliminate the label distribution deviation. Specifically, we first train the feature extractor on all domains to get domain invariant features, and then select an anchor domain to train the gaze regressor. We predict the gaze label on remaining domains and use a mapping function to align the labels. Finally, these aligned labels can be used to train gaze estimation models. Therefore, our method can be combined with any existing method. Experimental results show that our GLA method can effectively alleviate the label distribution shift, and SOTA gaze estimation methods can be further improved obviously.

Figures (6)

• Figure 1: The red scatters in (a) and (c) mean the labels of source domain $A$ and the corresponding blue scatters are the labels of source domain $B$. (b) and (d) are the scatter plots of estimated label (X-axis) and ground truth label (Y-axis).
• Figure 2: The schematic diagram of gaze direction. Gaze direction is originated from the gaze origin $\boldsymbol{o}$ and intersects with the screen at gaze target $\boldsymbol{t}_c$.
• Figure 3: Simulation analysis of the impact of gaze label deviation. (a) is the gaze deviation caused by each single variable of calibration error. (b) is the gaze deviation of coupling multiple calibration error variables.
• Figure 4: The examples of gaze deviation between gaze datasets. The red and blue arrows are truth gaze labels. In each column, two images from different datasets have similar head poses and eye rotation angles but significantly different gaze directions.
• Figure 5: Overview of the proposed GLA method. The GLA is the procedure before training the gaze estimation model and consists of six steps.