Iterative Pruning-based Model Compression for Pose Estimation on Resource-constrained Devices

Wonje Choi, Honguk Woo

Abstract: In this work, we propose a pruning-based model compression scheme, aiming at achieving an efficient model that has strength in both accuracy and inference time on an embedded device environment with limited resources. The proposed scheme consists of (1) pruning profiling and (2) iterative pruning via knowledge distillation. With the scheme, we develop a resource-efficient 2D pose estimation model using HRNet and evaluate the model on NVIDA JetsonNano with the Microsoft COCO keypoint dataset. Specifically, our compressed model obtains the fast pose estimation of 20.3 FPS on NVIDA JetsonNano, while maintaining a high accuracy of 74.1 AP. Compared to the conventional HRNet model without compression, the proposed compression technique achieves 33 % improvement in FPS with only 0.4 % degradation in AP.