Full metadata record
|dc.description.abstract||Adaptation of deep convolutional neural network has made revolutionary progress in human pose estimation and various applications in recent years have drawn considerable attention. However, prediction and localization of the keypoints in single and multi-person images are challenging problem. Towards this purpose we present a bottom-up box-free approach for the task of pose estimation and action recognition. We proposed StrongPose system model which uses part-based modeling to tackle object-part associations. The model utilizes a convolution network that learns how to detect Strong Keypoints Heat Maps (SKHM) and predict their comparative displacements, enabling us to group keypoints into person pose instances. Further, we produce Body Heat Maps (BHM) with the help of keypints which allows us to localize human body in the picture. The StrongPose framework is based on a fully-convolutional engineering and permits proficient inference, with runtime basically autonomous of the number of individuals display within the scene. Train and test on COCO data alone, our framework achieves COCO test-dev keypoint average precision of 0.708 using ResNet-101 and 0.725 using ResNet-152, which considerably outperforms all prior bottom-up pose estimation frameworks.||-|
|dc.title||StrongPose: Bottom-up, Strong Keypoints Heat Maps Base Person Pose Estimation||-|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.