Fine-tuned Domain-adapted Infant Pose (FiDIP)

Codes and experiments for the following paper:

Xiaofei Huang, Nihang Fu, Shuangjun Liu, Sarah Ostadabbas, “Invariant representation learning for infant pose estimation withsmall data,” in IEEE International Conference on Automatic Face and Gesture Recognition (FG), Dec. 2021

Contact:

Xiaofei Huang

Sarah Ostadabbas

Table of Contents

• Introduction
• What's New 🔥
• Main Results
• Environment
• Quick Start
• Synthetic Infant Data Generation
• Citation
• Acknowledgments

Introduction

This is an official pytorch implementation of Invariant Representation Learning for Infant Pose Estimation with Small Data. This work proposes a fine-tuned domain-adapted infant pose (FiDIP) estimation model, that transfers the knowledge of adult poses into estimating infant pose with the supervision of a domain adaptation technique on our released synthetic and real infant pose (SyRIP) dataset. Integrated with pose estimation backbone networks with varying complexity, FiDIP performs consistently better than the fine-tuned versions of those models. One of our best infant pose estimation performers on the state-of-the-art DarkPose model shows mean average precision (mAP) of 93.6. And the implementation of synthetic infant data generation is located under the root path.

What's New 🔥

• Release expanded SyRIP dataset including extra 400 real images in training set. Please contact us for the password of our dataset.
• Release new FiDIP model trained on the expanded SyRIP dataset.

Main Results

Performance comparison of three SOTA pose estimation models (SimpleBaseline, DarkPose, Pose-MobileNet) fine-tuned on MINI-RGBD, SyRIP-syn (synthesized data only) and SyRIP whole set and tested on SyRIP Test100.

Train Set	Method	Backbone	Input size	AP	AP50	AP75	AR	AR50	AR75
MINI-RGBD	SimpleBaseline	ResNet-50	384x288	69.2	95.8	78.0	72.4	97.0	81.0
SyRIP-syn	SimpleBaseline	ResNet-50	384x288	85.3	97.1	91.8	87.4	98.0	93.0
SyRIP	SimpleBaseline	ResNet-50	384x288	90.1	98.5	97.2	91.6	99.0	98.0
MINI-RGBD	DrakPose	HRNet-W48	384x288	85.2	98.6	95.3	87.0	99.0	96..0
SyRIP-syn	DrakPose	HRNet-W48	384x288	91.4	98.5	98.5	92.7	99.0	99.0
SyRIP	DrakPose	HRNet-W48	384x288	92.7	98.5	98.5	93.9	99.0	99.0
MINI-RGBD	Pose-MobileNet	MobileNetV2	224x224	12.3	38.1	3.8	21.6	52.0	14.0
SyRIP-syn	Pose-MobileNet	MobileNetV2	224x224	60.3	91.1	62.7	68.4	95.0	72.0
SyRIP	Pose-MobileNet	MobileNetV2	224x224	78.9	97.2	90.6	84.2	98.0	94.0

Generality of our FiDIP method to three models on the SyRIP Test100.

Method	Backbone	Input size	# Params	GFLOPs	AP	AP50	AP75	AR	AR50	AR75
SimpleBaseline	ResNet-50	384x288	32.42M	20.23	82.4	98.9	92.2	83.8	99.0	93.0
SimpleBaseline + Finetune	ResNet-50	384x288	32.42M	20.23	90.1	98.5	97.2	91.6	99.0	98.0
SimpleBaseline + FiDIP	ResNet-50	384x288	32.42M	20.23	91.1	98.5	98.5	92.6	99.0	99.0
DarkPose	HRNet-W48	384x288	60.65M	32.88	88.5	98.5	98.5	90.0	99.0	99.0
DarkPose + Finetune	HRNet-W48	384x288	60.65M	32.88	92.7	98.5	98.5	93.9	99.0	99.0
DarkPose + FiDIP	HRNet-W48	384x288	60.65M	32.88	93.6	98.5	98.5	94.6	99.0	99.0
Pose-MobileNet	MobileNetV2	224x224	3.91M	0.46	46.5	85.7	45.6	56.2	89.0	59.0
Pose-MobileNet + Finetune	MobileNetV2	224x224	3.91M	0.46	78.9	97.2	90.6	84.2	98.0	94.0
Pose-MobileNet + FiDIP	MobileNetV2	224x224	3.91M	0.46	79.2	99.0	89.4	84.1	99.0	92.0

Environment

The code is developed using python 3.12 with CUDA 12.1 on Ubuntu 18.04. NVIDIA GPUs are needed. The code is developed and tested using one NVIDIA TITAN Xp GPU card. Other platforms or GPU cards are not fully tested.

Quick Start

Installation

1. Clone this repo, and we'll call the directory that you cloned as ${POSE_ROOT} and get the following directory.

   ${POSE_ROOT}
   ├── data
   ├── experiments
   ├── lib
   ├── log
   ├── models
   ├── output
   ├── tools
   ├── README.md
   └── syn_generation
   
   

2. conda env create -f fidip_env.yml
   

3. Make libs:

   cd ${POSE_ROOT}/lib
   make
   

Download pretrained models

(1) download DarkPose pretrained models from TGA_models. (2) download our FiDIP pretrained model from FiDIP_models. Please download them under ${POSE_ROOT}/models, and make them look like this:

${POSE_ROOT}
 `-- models
     |-- hrnet_fidip.pth
     |-- mobile_fidip.pth
     `-- coco
         `-- w48_384x288.pth

Data preparation

For SyRIP data, please download from SyRIP dataset. The original SyRIP data can be download from `SyRIP.zip', and the data for FiDIP model training can be download from 'syrip_for_train.zip'. Download and extract them under {POSE_ROOT}/data, and make them look like this:

${POSE_ROOT}
|-- data
`-- |-- syrip
    `-- |-- annotations
        |   |-- person_keypoints_train_pre_infant.json
        |   |-- person_keypoints_train_infant.json   
        |   `-- person_keypoints_validate_infant.json
        `-- images
            |-- train_pre_infant
            |   |-- train00001.jpg
            |   |-- train00002.jpg
            |   |-- train00003.jpg
            |   |-- ... 
            |-- train_infant
            |   |-- train00001.jpg
            |   |-- train00002.jpg
            |   |-- train00003.jpg
            |   |-- ...  
            |   |-- train10001.jpg
            |   |-- train10002.jpg
            |   |-- train10003.jpg
            |   |-- ...  
            `-- validate_infant
                |-- test0.jpg
                |-- test1.jpg
                |-- test2.jpg
                |-- ... 

Note: our train_pre_infant dataset with only real/synthetic labels consists of 1904 samples from COCO Val2017 dataset and 2000 synthetic adult images from SURREAL dataset. If you need it to train model, please contact us. Or you can create your own train_pre_infant dataset.

Test on SyRIP validate dataset using FiDIP pretrained models

python tools/test_adaptive_model.py \
    --cfg experiments/coco/hrnet/w48_384x288_adam_lr1e-3_infant.yaml \
    TEST.MODEL_FILE models/hrnet_fidip.pth TEST.USE_GT_BBOX True

Training on SyRIP dataset

python tools/train_adaptive_model_hrnet.py \
    --cfg experiments/coco/hrnet/w48_384x288_adam_lr1e-3_infant.yaml

Synthetic Infant Data Generation

Please follow the README.md in the folder syn_generation.

Citation

If you use our code or models in your research, please cite with:

@inproceedings{huang2021infant,
  title={Invariant Representation Learning for Infant Pose Estimation with Small Data},
  author={Huang, Xiaofei and Fu, Nihang and Liu, Shuangjun and Ostadabbas, Sarah},
  booktitle={IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2021},
  month     = {December},
  year      = {2021}
}

Acknowledgement

Thanks for the open-source DarkPose

• Distribution-Aware Coordinate Representation for Human Pose Estimation, Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, Ce Zhu

Thanks for the SURREAL dataset

• Learning from Synthetic Humans, Gül Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev and Cordelia Schmid

License

• This code is for non-commertial purpose only.
• For further inquiry please contact: Augmented Cognition Lab at Northeastern University: http://www.northeastern.edu/ostadabbas/