In-Bed Pose Estimation: Deep Learning with Shallow Dataset

摘要: Although human pose estimation for various computer vision (CV) applications has been studied extensively in the last few decades, yet in-bed using camera-based methods ignored by CV community because it is assumed to be identical general purpose methods. However, its own specialized aspects and comes with specific challenges including notable differences lighting conditions throughout a day also having different distribution from common surveillance viewpoint. In this paper, we demonstrate that these significantly lessen effectiveness of existing models. order address variation challenge, infrared selective (IRS) image acquisition technique proposed provide uniform quality data under conditions. addition, deal unconventional perspective, 2-end histogram oriented gradient (HOG) rectification method presented. work, explored idea employing pre-trained convolutional neural network (CNN) model trained on large public datasets poses fine-tuning our shallow IRS dataset. We developed an imaging system collected several realistic life-size mannequins simulated hospital room environment. A CNN called machine (CPM) was repurposed intermediate layers. Using HOG method, performance CPM improved 26.4% PCK0.1 criteria compared without such rectification.