Adversarial Pose Estimation

A PyTorch implementation of the paper `Adversarial PoseNet- A Structure-aware Convolutional Network for Human Pose Estimation`. Course project for Computer Vision course at IIT Bombay

This repository aims to replicate the results of this paper. The idea is to augment the human pose estimation by using a GAN-based framework, where the (conditional) generator learns the distribution P(y x), where x is the image and y is the heatmap for the person. Typical keypoint detectors simply employ a similarity based loss (MSE or cross-entropy) on the predicted heatmaps with the ground-truth heatmaps. However, these losses can predicted smooth outputs as they are averaged over the entire spatial domain. The idea here is to make the predictions ‘‘crisper and sharper’’ by employing discriminators that differentiate between ground-truth and predicted heatmaps in 2 different ways.

Github source code is here.