This repository implements pose estimation methods in PyTorch.

The file lsp_mpii.h5 contains the annotations of MPII, LSP training data and LSP test data.
Place LSP, MPII images in data/LSP/images and data/mpii/images.
Place coco annotations in data/coco/annotations and images in data/coco/images, as suggested by cocoapi. The file valid_id contains the image_ids used for validation.

Compile the C implementation of the associative embedding loss. Code credit umich-vl/pose-ae-train.

cd src/extensions/AE
python build.py  # be sure to have visible cuda device

• data: put the training / testing data here
• src:
  • models: model definition
  • datasets: dataset definition
  • extensions:
    • AE: code from Associative Embedding.
      torch.autograd.Function extension for computing the loss.
  • utils

All the other folders represents different tasks. Each contains a training script train.py and definition of command-line options opts.py.

• hgpose: training code for Stacked Hourglass Networks for Human Pose Estimation. Single-Person.

• hgpose-ae: training code for Associative Embedding: End-to-end Learning for Joint Detection and Grouping. Multi-Person.

  COCO test compare, test on the images in valid_id

  • Author's pretrained checkpoint. 362 epochs with batch_size 32, 1000 iters per epoch and lr decay from 2e-4 to 1e-5 at 200000 iters.
  • My training result at 150 epochs with batch_size 16, 2465 iters per epoch and consistent lr of 2e-4. Roughly equals to half the progress of author's pretrained checkpoint.

• advpose: training code for Self Adversarial Training for Human Pose Estimation. Single-Person.

  • Comparison of training accuracy over steps
  

• advpose-ae: training code combining advpose with AE_loss. Multi-Person.

• advpose-ae: Only supports single gpu. Multi-gpu training get stucked randomly. The problem seems to be caused by the AE_loss extension.

• [ ] visualization
• [ ] example of usage