This repository contains a Keras implementation of the Pose Residual Network (PRN) presented in our ECCV 2018 paper:
Muhammed Kocabas, Salih Karagoz, Emre Akbas. MultiPoseNet: Fast Multi-Person Pose Estimation using Pose Residual Network. In ECCV, 2018. Arxiv
PRN is described in Section 3.2 of the paper.
We have tested our method on COCO Dataset
python
tensorflow
keras
numpy
tqdm
pycocotools
progress
scikit-image
-
Clone this repository:
git clone https://github.com/mkocabas/pose-residual-network.git
-
Install Tensorflow.
-
pip install -r src/requirements.txt
-
To download COCO dataset train2017 and val2017 annotations run:
bash data/coco.sh
. (data size: ~240Mb)
python main.py
For more options take a look at opt.py
Results on COCO val2017 Ground Truth data.
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets= 20 ] = 0.894
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets= 20 ] = 0.971
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets= 20 ] = 0.912
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.875
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.918
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 20 ] = 0.909
Average Recall (AR) @[ IoU=0.50 | area= all | maxDets= 20 ] = 0.972
Average Recall (AR) @[ IoU=0.75 | area= all | maxDets= 20 ] = 0.928
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.896
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.947
If you find this code useful for your research, please consider citing our paper:
@Inproceedings{kocabas18prn,
Title = {Multi{P}ose{N}et: Fast Multi-Person Pose Estimation using Pose Residual Network},
Author = {Kocabas, Muhammed and Karagoz, Salih and Akbas, Emre},
Booktitle = {European Conference on Computer Vision (ECCV)},
Year = {2018}
}