协调前景失衡的坐标注意力人群计数网络

rowd counting aims to estimate the total number of people accurately in the image and present their distribution. The images in the relevant dataset usually involve various scenes and a lot of people. Even the most advanced counting models are unable to focus heads clearly and accurately in extremely dense scenes due to the lack of detail in the head features and the heavy occlusion. In addition, there are a lot of background noises in complex scenes that can be easily misjudged, which will cause confusion to the recognition of people and also affect the counting results. In order to improve the representation ability of human head, the coordinate attention module is used to find key information, which considers the importance of structure information for feature extraction by encoding features from three dimensions, and adjusts the weights of different features adaptively to further emphasize regions of interest while suppressing background interference. Furthermore, in order to obtain higher quality density estimation maps, we proposed the foreground imbalance loss to adjust the importance of the foreground and background predicted errors, forcing the predictions to concentrate around the annotated point and suppressing background features. Experiments on multiple datasets show that the proposed method achieves better or competitive results and obtains clearer and more accurate density estimation maps compared to state-of-the-art methods.