A Simple Yet Effective Data Augmentation for Human Pose Estimation
Tien-Dat Tran, Xuan-Thuy Vo, G. F. Cao, Kang-Hyun Jo · 2023
Accurate occluded key point identification is a challenge and hot topic for human pose estimation. To make the occluded or invisible keypoint better, data augmentation play an important role which makes the network overcome complex case. In this paper, we want to apply the mosaic and mix-up technique which is a powerful method to tackle the problem. Furthermore, data augmentation demonstrates its superiority over other methods without enlarging the computational cost. Correspondingly, the proposed work focuses on powerful data augmentation for occluded keypoints. First, following a human detection in the detector network, feed the human proposal region into the data augmentation, which makes the network can learn more about the occluded cases. The data after data augmentation then apply to train for the pose estimator. The estimator collects more information in occluded keypoints, illustrating higher precision efficiency. The outputs of our experiments would also demonstrate a distinction between the use of mix-up and mosaic data augmentation and existing approaches. The predicted joint heatmaps are more accurate than the baseline technique despite using the same amount of parameters due to the transition to a high-resolution network (HRNet) for the pose estimator. Regarding AP, the proposed design outperforms the baseline network which is HRNet by 1.0 points, but in the occluded case, the pose estimator performs much better. Additionally, the COCO 2017 benchmarks, now accessible as an open and the most popular dataset for pose estimation, were used to train the proposed network.