An Efficient Approach Using Knowledge Distillation Methods to Stabilize Performance in a Lightweight Top-Down Posture Estimation Network.

Changhyun Park,Hean Sung Lee,Jaeho Lee,Sangyoun Lee,Woo Jin Kim,Han Byeol Bae

doi:10.3390/s21227640

Changhyun Park, Hean Sung Lee + Show 4 more

Open Access

https://doi.org/10.3390/s21227640

Copy DOI

Abstract

Multi-person pose estimation has been gaining considerable interest due to its use in several real-world applications, such as activity recognition, motion capture, and augmented reality. Although the improvement of the accuracy and speed of multi-person pose estimation techniques has been recently studied, limitations still exist in balancing these two aspects. In this paper, a novel knowledge distilled lightweight top-down pose network (KDLPN) is proposed that balances computational complexity and accuracy. For the first time in multi-person pose estimation, a network that reduces computational complexity by applying a “Pelee” structure and shuffles pixels in the dense upsampling convolution layer to reduce the number of channels is presented. Furthermore, to prevent performance degradation because of the reduced computational complexity, knowledge distillation is applied to establish the pose estimation network as a teacher network. The method performance is evaluated on the MSCOCO dataset. Experimental results demonstrate that our KDLPN network significantly reduces 95% of the parameters required by state-of-the-art methods with minimal performance degradation. Moreover, our method is compared with other pose estimation methods to substantiate the importance of computational complexity reduction and its effectiveness.

Highlights

The demand for human pose estimation has increased over time as it is essential for detecting human behaviors and for numerous applications such as human-computer interaction [1], human action recognition [2], and human performance analysis [3]
Inspired by PeleeNet [24], we propose knowledge distilled lightweight top-down pose network (KDLPN), a network that minimizes computational complexity while shuffling the pixels in the decoder previously introduced in the dense upsampling convolution (DUC) layer [25]
We compared our methods to other current state-of-the-art top-down-based human pose estimation methods such as Regional multiperson pose estimation (RMPE), Mask-RCNN [57], and G-RMI [19]

Summary

Introduction

The demand for human pose estimation has increased over time as it is essential for detecting human behaviors and for numerous applications such as human-computer interaction [1], human action recognition [2], and human performance analysis [3]. Human pose estimation has been studied as a close-up technique requiring a balance between accuracy and low computational complexity. Traditional approaches such as histogram of oriented gradient (HOG) [4] and Edgelet [5] extract discriminative features from images and assign a class to the feature vector. They cannot adequately determine the accurate location of body parts in a human figure [6]. Owing to the feature extraction capabilities of CNNs, the research paradigm of human pose estimation shifted from classic approaches to deep learning [7,8,9]. I.e., bottom-up and top-down approaches, of deep-learning-based methods, have been employed to overcome the limitations of handcrafting-based methods during the transition

Methods

Results

Conclusion