Robust Pose Recognition Using Deep Learning
Current pose estimation methods make unrealistic assumptions regarding the body postures. Here, we seek to propose a general scheme which does not make assumptions regarding the relative position of body parts. Practitioners of Indian classical dances suc
- PDF / 802,241 Bytes
- 13 Pages / 439.37 x 666.142 pts Page_size
- 117 Downloads / 277 Views
Abstract Current pose estimation methods make unrealistic assumptions regarding the body postures. Here, we seek to propose a general scheme which does not make assumptions regarding the relative position of body parts. Practitioners of Indian classical dances such as Bharatnatyam often enact several dramatic postures called Karanas. However, several challenges such as long flowing dresses of dancers, occlusions, change of camera viewpoint, poor lighting etc. affect the performance of stateof-the-art pose estimation algorithms [1, 2] adversely. Body postures enacted by practitioners performing Yoga also violate the assumptions used in current techniques for estimating pose. In this work, we adopt an image recognition approach to tackle this problem. We propose a dataset consisting of 864 images of 12 Karanas captured under controlled laboratory conditions and 1260 real-world images of 14 Karanas obtained from Youtube videos for Bharatnatyam. We also created a new dataset consisting of 400 real-world images of 8 Yoga postures. We use two deep
A. Mohanty (✉) ⋅ R.R. Sahay Department of Electrical Engineering, Indian Institute of Technology Kharagpur, Kharagpur, India e-mail: [email protected] R.R. Sahay e-mail: [email protected] A. Ahmed ⋅ T. Goswami ⋅ A. Das Department of Computer Science and Technology, Indian Institute of Engineering Science and Technology, Shibpur, Kolkata, India e-mail: [email protected] T. Goswami e-mail: [email protected] A. Das e-mail: [email protected] P. Vaishnavi Sardar Vallabhai National Institute of Technology Surat, Surat, India e-mail: [email protected] © Springer Science+Business Media Singapore 2017 B. Raman et al. (eds.), Proceedings of International Conference on Computer Vision and Image Processing, Advances in Intelligent Systems and Computing 460, DOI 10.1007/978-981-10-2107-7_9
93
94
A. Mohanty et al.
learning methodologies, namely, convolutional neural network (CNN) and stacked auto encoder (SAE) and demonstrate that both these techniques achieve high recognition rates on the proposed datasets. Keywords Pose estimation Stacked auto encoder
⋅
Deep learning
⋅
Convolutional neural network
⋅
1 Introduction The current state-of-the-art pose estimation methods are not flexible enough to model horizontal people, suffers from double counting phenomena (when both left and right legs lie on same image region) and gets confused when objects partially occlude people. Earlier works on pose estimation [1, 2], impose a stick-man model on the image of the body and assume that the head lies above the torso. Similarly, shoulder joints are supposed to be higher than the hip joint and legs. However, these assumptions are unrealistic and are violated under typical scenarios shown in this work. As an example, we show how one state-of-the-art approach [1] fails to estimate the pose correctly for an image taken from the standard PARSE dataset as shown in Fig. 1a, b. The images of Indian classical dance (ICD) and Yoga too have such complex configuration of body postures where c
Data Loading...
