Pytorch implementation of pose proposal network
We train the network on MPII human pose dataset but don't evaluate the performance. I don't think it can match the performance mentioned in paper due to some problems(see below).
- pytorch 0.4.1
The speed depend on numbers of person. More persons there are, a little slower it runs.
The above figure is iou loss. you can see it that decrease not very better than other losses(under images folder). In fact, when I print it and ground truth, I find relative large error between them, especially when ground truth iou is low. For example while ground truth iou is 0.2, prediction iou is 0.5 or 0.6. So at parse stage, I just use resp instead of resp * iou.
Another problem is that I can't figure out one formula: S * E * D in the paper. I just use D * E * D.
First you need to put MPII dataset images in data/images. Then You can train from scratch or download our well trained weights.
- train from scratch: download pretrained weights resnet18 weights and resnet50 weights, and put them in src/model/pretrain_weight directory.
- well trained weights: download resnet18 weights and resnet50 weights and put them in src/checkpoint directory.
when I do this project, I mainly refer to this repo: https://github.com/hizhangp/yolo_tensorflow.