| Autor: |
Kumar, Pranjal, Chauhan, Siddhartha |
| Zdroj: |
Evolving Systems; Apr2024, Vol. 15 Issue 2, p659-667, 9p |
| Abstrakt: |
A challenging problem for robotic interaction and augmented reality is the estimation and tracking of human poses in images and videos. Pose estimation using deep neural networks has shown encouraging results in recent approaches. The environmental sensitivity and computational complexity of conventional pose estimation methods are major drawbacks. In light of these issues, this paper proposes a novel approach that uses DenseNet and CNN-based transfer learning to learn by explicitly exploiting the skeletal data. Other imageNet pre-trained models along with probabilistic and regression losses are used for comparative study. A widely accepted benchmark pose estimation dataset, FLIC (Frames Labelled in Cinema) serves as the basis for our evaluation and comparison. As a result of our experiments with an R 2 score of 0.948, we recommend probabilistic loss over regression loss as the new baseline for future downstream tasks and fine-tuning-based transfer learning techniques for pose estimation. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
Full text from SpringerLink