Towards Adapting CLIP for Gaze Object Prediction

This paper aims to investigate the problem of gaze object prediction in single images. We propose an application-friendly network based on CLIP for gaze object prediction. To avoid domain bias, we utilize a shallow feature adapter that transfers pre-trained features to target-oriented ones. Secondly, we introduce a pooling attention block to exploit the joint representation of multimodal elements, reducing gaze point deviation. Additionally, we introduce a loss that measures the prediction quality by comparing the distribution difference between the model's predictions heatmaps and the ground truth. Extensive experiments demonstrate the superior performance of our model compared to previous models. We will provide the method code at: https://github.com/fadaishaitaiyang/CCLIP.git.