In this paper, we propose two additional heatmap constraint methods that can be integrated into existing detection-based 3D Hand Pose Estimation backbone networks. Our methods effectively reduce the gap between the training and inferencing, which is often caused by the post processing in detection-based methods, by proposing a constraint on the intermediate dense representation-the heatmap. Our methods achieve excellent results on both the NYU and ICVL datasets, demonstrating the effectiveness of our proposed method.