Spatial probabilistic characterization techniques
Gaze-LLE outputs a spatial heat map normalized to the [0,1] interval, with each pixel value corresponding to the gaze probability density. This characterization is more explanatory than traditional coordinate regression, and is able to reflect both the gaze focus region and the uncertainty level. For the technical implementation, the model uses a lightweight decoder to convert the visual features extracted by DINOv2 into a probability distribution map with 256×256 resolution.
The output format is particularly suitable for multiplayer scene analysis, where a single forward propagation can be used to generate a heatmap of the gaze of all individuals in the scene. In the user interface design, gaze regions can be generated by probabilistic threshold filtering, or the original map can be overlaid for visual analysis. Experiments show that the method achieves an AUC metric of 92.3% on the GazeFollow dataset.
This answer comes from the articleGaze-LLE: A Target Prediction Tool for Character Gaze in VideoThe
