Semi-supervised Crowd Counting in Educational Scenarios

In the field of education, a significant amount of crowd image data is generated on campus every day. However, the use of fully supervised learning requires substantial annotation costs and time, which is obviously unrealistic. To alleviate the heavy annotation burden when training a reliable crowd counting model and to enable the model to benefit from more data, thereby becoming more practical and accurate, this paper proposes a new semi-supervised method based on the "Mean-Teacher" framework. The framework employs VGG16 for feature extraction and incorporates a multi-scale feature fusion module to adapt to variations in image scale and crowd density. Innovations include the design of a fine-grained spatial reconstruction module (aimed at refining features), an attention module (aimed at suppressing background interference), and the application of a semi-supervised learning framework in the educational context. Experiments indicate that this method performs effectively on both open-access datasets and the self-collected educational datasets. Furthermore, the ablation study highlights the contribution of two modules proposed as well as the effectiveness of the semi-supervised method employed.