Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose

This repository contains training code for the paper Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose. This work heavily optimizes the OpenPose approach to reach real-time inference on CPU with negliable accuracy drop. It detects a skeleton (which consists of keypoints and connections between them) to identify human poses for every person inside the image. The pose may contain up to 18 keypoints: ears, eyes, nose, neck, shoulders, elbows, wrists, hips, knees, and ankles. On COCO 2017 Keypoint Detection validation set this code achives 40% AP for the single scale inference (no flip or any post-processing done). The result can be reproduced using this repository. This repo significantly overlaps with https://github.com/opencv/openvino_training_extensions, however contains just the necessary code for human pose estimation.

:fire: Check out our new work on accurate (and still fast) single-person pose estimation, which ranked 10^th on CVPR'19 Look-Into-Person challenge.

:fire::fire: Check out our lightweight 3D pose estimation, which is based on Single-Shot Multi-Person 3D Pose Estimation From Monocular RGB paper and this work.

Table of Contents

• Requirements
• Prerequisites
• Training
• Validation
• Pre-trained model
• C++ demo
• Python demo
• Citation

Other Implementations

• TensorFlow by murdockhou.
• OpenVINO by Pavel Druzhkov.

Requirements

• Ubuntu 16.04
• Python 3.6
• PyTorch 0.4.1 (should also work with 1.0, but not tested)

Prerequisites

1. Download COCO 2017 dataset: http://cocodataset.org/#download (train, val, annotations) and unpack it to <COCO_HOME> folder.
2. Install requirements pip install -r requirements.txt

Training

Training consists of 3 steps (given AP values for full validation dataset):

• Training from MobileNet weights. Expected AP after this step is ~38%.
• Training from weights, obtained from previous step. Expected AP after this step is ~39%.
• Training from weights, obtained from previous step and increased number of refinement stages to 3 in network. Expected AP after this step is ~40% (for the network with 1 refinement stage, two next are discarded).

1. Download pre-trained MobileNet v1 weights mobilenet_sgd_68.848.pth.tar from: https://github.com/marvis/pytorch-mobilenet (sgd option). If this doesn't work, download from GoogleDrive.

2. Convert train annotations in internal format. Run python scripts/prepare_train_labels.py --labels <COCO_HOME>/annotations/person_keypoints_train2017.json. It will produce prepared_train_annotation.pkl with converted in internal format annotations.

   [OPTIONAL] For fast validation it is recommended to make subset of validation dataset. Run python scripts/make_val_subset.py --labels <COCO_HOME>/annotations/person_keypoints_val2017.json. It will produce val_subset.json with annotations just for 250 random images (out of 5000).

3. To train from MobileNet weights, run python train.py --train-images-folder <COCO_HOME>/train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder <COCO_HOME>/val2017/ --checkpoint-path <path_to>/mobilenet_sgd_68.848.pth.tar --from-mobilenet

4. Next, to train from checkpoint from previous step, run python train.py --train-images-folder <COCO_HOME>/train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder <COCO_HOME>/val2017/ --checkpoint-path <path_to>/checkpoint_iter_420000.pth --weights-only

5. Finally, to train from checkpoint from previous step and 3 refinement stages in network, run python train.py --train-images-folder <COCO_HOME>/train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder <COCO_HOME>/val2017/ --checkpoint-path <path_to>/checkpoint_iter_280000.pth --weights-only --num-refinement-stages 3. We took checkpoint after 370000 iterations as the final one.

We did not perform the best checkpoint selection at any step, so similar result may be achieved after less number of iterations.

Known issue

We observe this error with maximum number of open files (ulimit -n) equals to 1024:

  File "train.py", line 164, in <module>
    args.log_after, args.val_labels, args.val_images_folder, args.val_output_name, args.checkpoint_after, args.val_after)
  File "train.py", line 77, in train
    for _, batch_data in enumerate(train_loader):
  File "/<path>/python3.6/site-packages/torch/utils/data/dataloader.py", line 330, in __next__
    idx, batch = self._get_batch()
  File "/<path>/python3.6/site-packages/torch/utils/data/dataloader.py", line 309, in _get_batch
    return self.data_queue.get()
  File "/<path>/python3.6/multiprocessing/queues.py", line 337, in get
    return _ForkingPickler.loads(res)
  File "/<path>/python3.6/site-packages/torch/multiprocessing/reductions.py", line 151, in rebuild_storage_fd
    fd = df.detach()
  File "/<path>/python3.6/multiprocessing/resource_sharer.py", line 58, in detach
    return reduction.recv_handle(conn)
  File "/<path>/python3.6/multiprocessing/reduction.py", line 182, in recv_handle
    return recvfds(s, 1)[0]
  File "/<path>/python3.6/multiprocessing/reduction.py", line 161, in recvfds
    len(ancdata))
RuntimeError: received 0 items of ancdata

To get rid of it, increase the limit to bigger number, e.g. 65536, run in the terminal: ulimit -n 65536

Validation

1. Run python val.py --labels <COCO_HOME>/annotations/person_keypoints_val2017.json --images-folder <COCO_HOME>/val2017 --checkpoint-path <CHECKPOINT>

Pre-trained model

The model expects normalized image (mean=[128, 128, 128], scale=[1/256, 1/256, 1/256]) in planar BGR format. Pre-trained on COCO model is available at: https://download.01.org/opencv/openvino_training_extensions/models/human_pose_estimation/checkpoint_iter_370000.pth, it has 40% of AP on COCO validation set (38.6% of AP on the val subset).

Conversion to OpenVINO format

1. Convert PyTorch model to ONNX format: run script in terminal python scripts/convert_to_onnx.py --checkpoint-path <CHECKPOINT>. It produces human-pose-estimation.onnx.
2. Convert ONNX model to OpenVINO format with Model Optimizer: run in terminal python <OpenVINO_INSTALL_DIR>/deployment_tools/model_optimizer/mo.py --input_model human-pose-estimation.onnx --input data --mean_values data[128.0,128.0,128.0] --scale_values data[256] --output stage_1_output_0_pafs,stage_1_output_1_heatmaps. This produces model human-pose-estimation.xml and weights human-pose-estimation.bin in single-precision floating-point format (FP32).

C++ Demo

C++ demo can be found in the Intel® OpenVINO™ toolkit, the corresponding model is human-pose-estimation-0001. Please follow the official instruction to run it.

Python Demo

@inproceedings{osokin2018lightweight_openpose,
    author={Osokin, Daniil},
    title={Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose},
    booktitle = {arXiv preprint arXiv:1811.12004},
    year = {2018}
}