Home

Awesome

KAPAO (Keypoints and Poses as Objects)

Accepted to ECCV 2022

KAPAO is an efficient single-stage multi-person human pose estimation method that models keypoints and poses as objects within a dense anchor-based detection framework. KAPAO simultaneously detects pose objects and keypoint objects and fuses the detections to predict human poses:

alt text

When not using test-time augmentation (TTA), KAPAO is much faster and more accurate than previous single-stage methods like DEKR, HigherHRNet, HigherHRNet + SWAHR, and CenterGroup:

alt text

This repository contains the official PyTorch implementation for the paper: <br> Rethinking Keypoint Representations: Modeling Keypoints and Poses as Objects for Multi-Person Human Pose Estimation.

Our code was forked from ultralytics/yolov5 at commit 5487451.

Setup

  1. If you haven't already, install Anaconda or Miniconda.
  2. Create a new conda environment with Python 3.6: $ conda create -n kapao python=3.6.
  3. Activate the environment: $ conda activate kapao
  4. Clone this repo: $ git clone https://github.com/wmcnally/kapao.git
  5. Install the dependencies: $ cd kapao && pip install -r requirements.txt
  6. Download the trained models: $ python data/scripts/download_models.py

Inference Demos

Note: FPS calculations include all processing (i.e., including image loading, resizing, inference, plotting / tracking, etc.). See script arguments for inference options.


Static Image

To generate the four images in the GIF above:

  1. $ python demos/image.py --bbox
  2. $ python demos/image.py --bbox --pose --face --no-kp-dets
  3. $ python demos/image.py --bbox --pose --face --no-kp-dets --kp-bbox
  4. $ python demos/image.py --pose --face

Shuffling Video

KAPAO runs fastest on low resolution video with few people in the frame. This demo runs KAPAO-S on a single-person 480p dance video using an input size of 1024. The inference speed is ~9.5 FPS on our CPU, and ~60 FPS on our TITAN Xp.

CPU inference:<br> alt text<br>

To display the results in real-time: <br> $ python demos/video.py --face --display

To create the GIF above:<br> $ python demos/video.py --face --device cpu --gif

CPU specs:<br> Intel Core i7-8700K<br> 16GB DDR4 3000MHz<br> Samsung 970 Pro M.2 NVMe SSD<br>


Flash Mob Video

This demo runs KAPAO-S on a 720p flash mob video using an input size of 1280.

GPU inference:<br> alt text<br>

To display the results in real-time: <br> $ python demos/video.py --yt-id 2DiQUX11YaY --tag 136 --imgsz 1280 --color 255 0 255 --start 188 --end 196 --display

To create the GIF above:<br> $ python demos/video.py --yt-id 2DiQUX11YaY --tag 136 --imgsz 1280 --color 255 0 255 --start 188 --end 196 --gif


Red Light Green Light

This demo runs KAPAO-L on a 480p clip from the TV show Squid Game using an input size of 1024. The plotted poses constitute keypoint objects only.

GPU inference:<br> alt text<br>

To display the results in real-time:<br> $ python demos/video.py --yt-id nrchfeybHmw --imgsz 1024 --weights kapao_l_coco.pt --conf-thres-kp 0.01 --kp-obj --face --start 56 --end 72 --display

To create the GIF above:<br> $ python demos/video.py --yt-id nrchfeybHmw --imgsz 1024 --weights kapao_l_coco.pt --conf-thres-kp 0.01 --kp-obj --face --start 56 --end 72 --gif


Squash Video

This demo runs KAPAO-S on a 1080p slow motion squash video. It uses a simple player tracking algorithm based on the frame-to-frame pose differences.

GPU inference:<br> alt text<br>

To display the inference results in real-time: <br> $ python demos/squash.py --display --fps

To create the GIF above:<br> $ python demos/squash.py --start 42 --end 50 --gif --fps


Depth Video

Pose objects generalize well and can even be detected in depth video. Here KAPAO-S was run on a depth video from a fencing action recognition dataset.

alt text<br>

The depth video above can be downloaded directly from here. To create the GIF above:<br> $ python demos/video.py -p 2016-01-04_21-33-35_Depth.avi --face --start 0 --end -1 --gif --gif-size 480 360


Web Demo

A web demo was integrated to Huggingface Spaces with Gradio (credit to @AK391). It uses KAPAO-S to run CPU inference on short video clips.

COCO Experiments

Download the COCO dataset: $ sh data/scripts/get_coco_kp.sh

Validation (without TTA)

Validation (with TTA)

Testing

Training

The following commands were used to train the KAPAO models on 4 V100s with 32GB memory each.

KAPAO-S:

python -m torch.distributed.launch --nproc_per_node 4 train.py \
--img 1280 \
--batch 128 \
--epochs 500 \
--data data/coco-kp.yaml \
--hyp data/hyps/hyp.kp-p6.yaml \
--val-scales 1 \
--val-flips -1 \
--weights yolov5s6.pt \
--project runs/s_e500 \
--name train \
--workers 128

KAPAO-M:

python train.py \
--img 1280 \
--batch 72 \
--epochs 500 \
--data data/coco-kp.yaml \
--hyp data/hyps/hyp.kp-p6.yaml \
--val-scales 1 \
--val-flips -1 \
--weights yolov5m6.pt \
--project runs/m_e500 \
--name train \
--workers 128

KAPAO-L:

python train.py \
--img 1280 \
--batch 48 \
--epochs 500 \
--data data/coco-kp.yaml \
--hyp data/hyps/hyp.kp-p6.yaml \
--val-scales 1 \
--val-flips -1 \
--weights yolov5l6.pt \
--project runs/l_e500 \
--name train \
--workers 128

Note: DDP is usually recommended but we found training was less stable for KAPAO-M/L using DDP. We are investigating this issue.

CrowdPose Experiments

Testing

Training

The following commands were used to train the KAPAO models on 4 V100s with 32GB memory each. Training was performed on the trainval split with no validation. The test results above were generated using the last model checkpoint.

KAPAO-S:

python -m torch.distributed.launch --nproc_per_node 4 train.py \
--img 1280 \
--batch 128 \
--epochs 300 \
--data data/crowdpose.yaml \
--hyp data/hyps/hyp.kp-p6.yaml \
--val-scales 1 \
--val-flips -1 \
--weights yolov5s6.pt \
--project runs/cp_s_e300 \
--name train \
--workers 128 \
--noval

KAPAO-M:

python train.py \
--img 1280 \
--batch 72 \
--epochs 300 \
--data data/crowdpose.yaml \
--hyp data/hyps/hyp.kp-p6.yaml \
--val-scales 1 \
--val-flips -1 \
--weights yolov5m6.pt \
--project runs/cp_m_e300 \
--name train \
--workers 128 \
--noval

KAPAO-L:

python train.py \
--img 1280 \
--batch 48 \
--epochs 300 \
--data data/crowdpose.yaml \
--hyp data/hyps/hyp.kp-p6.yaml \
--val-scales 1 \
--val-flips -1 \
--weights yolov5l6.pt \
--project runs/cp_l_e300 \
--name train \
--workers 128 \
--noval

Acknowledgements

This work was supported in part by Compute Canada, the Canada Research Chairs Program, the Natural Sciences and Engineering Research Council of Canada, a Microsoft Azure Grant, and an NVIDIA Hardware Grant.

If you find this repo is helpful in your research, please cite our paper:

@article{mcnally2021kapao,
  title={Rethinking Keypoint Representations: Modeling Keypoints and Poses as Objects for Multi-Person Human Pose Estimation},
  author={McNally, William and Vats, Kanav and Wong, Alexander and McPhee, John},
  journal={arXiv preprint arXiv:2111.08557},
  year={2021}
}

Please also consider citing our previous works:

@inproceedings{mcnally2021deepdarts,
  title={DeepDarts: Modeling Keypoints as Objects for Automatic Scorekeeping in Darts using a Single Camera},
  author={McNally, William and Walters, Pascale and Vats, Kanav and Wong, Alexander and McPhee, John},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4547--4556},
  year={2021}
}

@article{mcnally2021evopose2d,
  title={EvoPose2D: Pushing the Boundaries of 2D Human Pose Estimation Using Accelerated Neuroevolution With Weight Transfer},
  author={McNally, William and Vats, Kanav and Wong, Alexander and McPhee, John},
  journal={IEEE Access},
  volume={9},
  pages={139403--139414},
  year={2021},
  publisher={IEEE}
}