An Investigation of 2D Keypoints Detection on Challenging Scenarios Using Depthwise Separable Convolutions: A Hand Pose Estimation Case Study

• Willams Costa UFPE
• Lucas Figueiredo UFPE
• João Marcelo Teixeira UFPE
• João Paulo Lima UFPE
• Veronica Teichrieb UFPE

Resumo

2D keypoints detection is a computer vision task applicable to several fields such as hand, face, and body tracking, which provides useful information for spatial analytics, gestural interactions, and augmented reality applications. This work investigates the usage of depthwise separable convolutions (an optimized convolution operation) to speed up the inference time on a largely used architecture for 2D keypoints estimation. We evaluate the impacts on the precision and performance of such optimization on a hand pose estimation task. We also extend the evaluation towards simulated challenging scenarios of defocused lens, motion blur, occlusions, and noisy images to understand how these stress situations affect both the original and the optimized architectures. We show that the execution time can be improved on average by 12.8% with an accuracy compromise of less than 1 pixel (mean EPE). The experiments on challenging scenarios revealed that the model powered by depthwise separable convolutions is most fit for the occlusion cases and noisy environments while suffering more on the motion blur simulated scenarios.

Palavras-chave: Graphics, Computer vision, Pose estimation, Computer architecture, Noise measurement, Task analysis, Stress