Simple Fine-Tuning Attention Modules for Human Pose Estimation

The convolution neural networks (CNNs) have achieved the best performance not only for human pose estimation but also for other computer vision tasks (e.g., object detection, semantic segmentation, image classification). Then this paper focuses on a useful attention module (AM) for feed-forward CNNs. Firstly, feed the feature map after a block in the backbone network into the attention module, split into two separate dimensions, channel and spatial. After that, the AM combines these two feature maps by multiplication and gives it to the next block in the backbone. The network can capture the information in the long-range dependencies (channel) and the spatial data, which can gain better performance in accuracy. Therefore, our experimental results will illustrate how different between when using the attention module and the existing methods. As a result, the predicted joint heatmap maintains the accuracy and spatially better with the simple baseline. Besides, the proposed architecture gains 1.0 points in AP higher than the baseline. Moreover, the proposed network trained on COCO 2017 benchmarks, which is an accessible dataset nowadays.