This repository contains 3D multi-person pose estimation demo in PyTorch. Intel OpenVINO™ backend can be used for fast inference on CPU. This demo is based on Lightweight OpenPose and Single-Shot Multi-Person 3D Pose Estimation From Monocular RGB papers. It detects 2D coordinates of up to 18 types of keypoints: ears, eyes, nose, neck, shoulders, elbows, wrists, hips, knees, and ankles, as well as their 3D coordinates. It was trained on MS COCO and CMU Panoptic datasets and achieves 100 mm MPJPE (mean per joint position error) on CMU Panoptic subset. This repository significantly overlaps with https://github.com/opencv/open_model_zoo/, however contains just the necessary code for 3D human pose estimation demo.
The major part of this work was done by Mariia Ageeva, when she was the 🔝🚀🔥 intern at Intel.
pip install -r requirements.txt
python setup.py build_ext
Pre-trained model is available at Google Drive.
To run the demo, pass path to the pre-trained checkpoint and camera id (or path to video file):
python demo.py --model human-pose-estimation-3d.pth --video 0
Camera can capture scene under different view angles, so for correct scene visualization, please pass camera extrinsics and focal length with
--fxoptions correspondingly (extrinsics sample format can be found in data folder). In case no camera parameters provided, demo will use the default ones.
To run with OpenVINO, it is necessary to convert checkpoint to OpenVINO format:
python scripts/convert_to_onnx.py --checkpoint-path human-pose-estimation-3d.pth
python <OpenVINO_INSTALL_DIR>/deployment_tools/model_optimizer/mo.py --input_model human-pose-estimation-3d.onnx --input=data --mean_values=data[128.0,128.0,128.0] --scale_values=data[255.0,255.0,255.0] --output=features,heatmaps,pafs
To run the demo with OpenVINO inference, pass
--use-openvino option and specify device to infer on:
python demo.py --model human-pose-estimation-3d.xml --device CPU --use-openvino --video 0
To run with TensorRT, it is necessary to install it properly. Please, follow the official guide, these steps work for me:
python -m pip install nvidia-pyindex pip install nvidia-tensorrt==184.108.40.206
Convert checkpoint to TensorRT format:
python scripts/convert_to_trt.py --checkpoint-path human-pose-estimation-3d.pth
TensorRT does not support dynamic network input size reshape. Make sure you have set proper network input height, width with
--widthoptions during conversion (if not, there will be no detections). Default values work for a usual video with 16:9 aspect ratio (1280x720, 1920x1080). You can check the network input size with
print(scaled_img.shape)in the demo.py
To run the demo with TensorRT inference, pass
python demo.py --model human-pose-estimation-3d-trt.pth --use-tensorrt --video 0
I have observed ~10x network inference speedup on RTX 2060 (in comparison with default PyTorch 1.6.0+cu101 inference).