The invention puts forward a line of sight estimation method based on a generative adversarial network. The method comprises the following main contents of generating a texture, generating real data and refining eyes. The method comprises the following processes that: firstly, automatically aligning a face image with the texture space of the horizontal direction and the vertical direction of a 3Dmodel; then, mapping a synthesis image to a true domain by an unpaired pixel level domain adaptive technology; thirdly, using the annotation and synthesis data of a line of sight direction to pre-train a line of sight direction estimator; and finally, in a whole mapping process, executing a refined network to keep a line of sight direction, and using a pre-training network to serve as a conversioncirculation constraint from synthesis to truth to synthesis. By use of the method, a novel adversarial training method is used, the rendered synthesis image is mapped to a vivid domain, and accurateline of sight estimation can be obtained on a practical image without using any piece of additional flag data from a true user. For the situations of extreme head gesture, blur, long distance and thelike, the method can generate line of sight estimation with robustness.