Deep Reinforcement Learning for Active Human Pose Estimation Supplementary Material

1 Model Architecture See Fig. 1 for a description of the Pose-DRL architecture. The underlying pose estimation networks, DMHS (Popa, Zanfir, and Sminchisescu 2017) and MubyNet (Zanfir et al. 2018), as well as our agent were implemented in Caffe (Jia et al. 2014) and MATLAB. For the Faster R-CNN detector (Ren et al. 2015) we used a publicly available Tensorflow (Abadi et al. 2016) implementation,1 with ResNet-101 (He et al. 2016) as base feature extractor.