PyTorch implementation of the Dec 2018 paper: https://arxiv.org/abs/1812.08008
Go through estimate-pose.ipynb for training and evaluation code on sample image, video.
Model has been trained on MS-COCO 2014 dataset on 368x368 and 184x184 resolutions (model-wts-368.ckpt, model-wts-184.ckpt).
Has additional PAF's from Shoulder->Wrist and Hip->Ankle for improved matching in crowded scenes.
You can change the threholds for PAF map o/p values, Heatmap threshold and part matching threshold in CONFIG.py (for more, less conf joint preds vs less, more confident).
Part Matching formulation uses Munkres for one-one least cost matching.
Joint Matching: (With the help of predicted Part Affinity Field vectors)
Network Architecture: (1st 10 layers from VGG-16 as backbone(F), 4 PAF stages(L), 2 Heatmap stages(S))
Training Image, Generated PAF's, Generated Joint Heatmaps: