Figure 1

Figure 1

Comparison of the conventional keypoint detection framework and ours. (a) In the conventional framework for using a heatmap, the convolution kernel extracts the image feature map and obtains a local receptive field. Subsequently, the local field is compared to the heatmap point by point. The probability of keypoint occurrence is highest near the annotated position (red area) and gradually decreases as it moves farther away (blue area). (b) Our CSHT-Net, proposed in this study, breaks the strong connection between the model feature map and the GT label heatmap, allowing the feature map to explore long-distance information. Additionally, our network has augmented the receptive field using our improved combinatorial convolution block.