Distilled RSN: Lightweight Pose Estimation Using Knowledge Distillation
Jiu Yi, Haoyuan Liu, Hiroshi Watanabe · 2024
Current research on pose estimation often implements repeated functional blocks in model design to improve detection accuracy. Such a strategy results in increased computational complexity and resource consumption, failing to meet real-time inference demand. We discover this efficiency problem by retraining single and multiple functional modules of RSN and then applying the same metrics for evaluation. The inference performance only improves a bit by introducing computational costs several times. To address this issue, we present a novel model simplification strategy, Distilled RSN, which adopts knowledge distillation to refine the redundant RSN blocks into a single, efficient module. Our experiments demonstrate that our method outperforms many lightweight pose approaches in the COCO keypoint dataset. Compared to the method that applies only the original single RSN module for pose inference, we improved accuracy by 1.6% by our strategy.