Cross Domain Facial Expression Recognition Benchmark

Implementation of papers:

Cross-Domain Facial Expression Recognition: A Unified Evaluation Benchmark and Adversarial Graph Learning
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)，2021.
Tianshui Chen*, Tao Pu*, Hefeng Wu, Yuan Xie, Lingbo Liu, Liang Lin.
Adversarial Graph Representation Adaptation for Cross-Domain Facial Expression Recognition
ACM International Conference on Multimedia (ACM MM), 2020. (Oral Presentation)
Yuan Xie, Tianshui Chen, Tao Pu, Hefeng Wu, Liang Lin.

Environment

Ubuntu 16.04 LTS, Python 3.5, PyTorch 1.3

Note: We also provide docker image for this project, click here. (Tag: py3-pytorch1.3-agra)

Datasets

To apply for the AFE, please complete the AFE Database User Agreement and submit it to [email protected] or [email protected].

Note:

The AFE Database Agreement needs to be signed by the faculty member at a university or college and sent it by email.
In order to comply with relevant regulations, you need to apply for the image data of the following data sets by yourself, including CK+, JAFFE, SFEW 2.0, FER2013, ExpW, RAF.

Pre-Train Model

You can download pre-train models in Baidu Drive (password: tzrf) and OneDrive.

Note: To replace backbone of each methods, you should modify and run getPreTrainedModel_ResNet.py (or getPreTrainedModel_MobileNet.py) in the folder where you want to use the method.

Usage

Before run these script files, you should download datasets and pre-train model, and run getPreTrainedModel_ResNet.py (or getPreTrainedModel_MobileNet.py).

Run ICID

cd ICID
bash Train.sh

Run DFA

cd DFA
bash Train.sh

Run LPL

cd LPL
bash Train.sh

Run DETN

cd DETN
bash TrainOnSourceDomain.sh     # Train Model On Source Domain
bash TransferToTargetDomain.sh  # Then, Transfer Model to Target Domain

Run FTDNN

cd FTDNN
bash Train.sh

Run ECAN

cd ECAN
bash TrainOnSourceDomain.sh     # Train Model On Source Domain
bash TransferToTargetDomain.sh  # Then, Transfer Model to Target Domain

Run CADA

cd CADA
bash TrainOnSourceDomain.sh     # Train Model On Source Domain
bash TransferToTargetDomain.sh  # Then, Transfer Model to Target Domain

Run SAFN

cd SAFN
bash TrainWithSAFN.sh

Run SWD

cd SWD
bash Train.sh

Run AGRA

cd AGRA
bash TrainOnSourceDomain.sh     # Train Model On Source Domain
bash TransferToTargetDomain.sh  # Then, Transfer Model to Target Domain

Result

Souce Domain: RAF

Methods	Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ICID	ResNet-50	74.42	50.70	48.85	53.70	69.54	59.44
DFA	ResNet-50	64.26	44.44	43.07	45.79	56.86	50.88
LPL	ResNet-50	74.42	53.05	48.85	55.89	66.90	59.82
DETN	ResNet-50	78.22	55.89	49.40	52.29	47.58	56.68
FTDNN	ResNet-50	79.07	52.11	47.48	55.98	67.72	60.47
ECAN	ResNet-50	79.77	57.28	52.29	56.46	47.37	58.63
CADA	ResNet-50	72.09	52.11	53.44	57.61	63.15	59.68
SAFN	ResNet-50	75.97	61.03	52.98	55.64	64.91	62.11
SWD	ResNet-50	75.19	54.93	52.06	55.84	68.35	61.27
Ours	ResNet-50	85.27	61.50	56.43	58.95	68.50	66.13

Methods	Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ICID	ResNet-18	67.44	48.83	47.02	53.00	68.52	56.96
DFA	ResNet-18	54.26	42.25	38.30	47.88	47.42	46.02
LPL	ResNet-18	72.87	53.99	49.31	53.61	68.35	59.63
DETN	ResNet-18	64.19	52.11	42.25	42.01	43.92	48.90
FTDNN	ResNet-18	76.74	50.23	49.54	53.28	68.08	59.57
ECAN	ResNet-18	66.51	52.11	48.21	50.76	48.73	53.26
CADA	ResNet-18	73.64	55.40	52.29	54.71	63.74	59.96
SAFN	ResNet-18	68.99	49.30	50.46	53.31	68.32	58.08
SWD	ResNet-18	72.09	53.52	49.31	53.70	65.85	58.89
Ours	ResNet-18	77.52	61.03	52.75	54.94	69.70	63.19

Methods	Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ICID	MobileNet V2	57.36	37.56	38.30	44.47	60.64	47.67
DFA	MobileNet V2	41.86	35.21	29.36	42.36	43.66	38.49
LPL	MobileNet V2	59.69	40.38	40.14	50.13	62.26	50.52
DETN	MobileNet V2	53.49	40.38	35.09	45.88	45.26	44.02
FTDNN	MobileNet V2	71.32	46.01	45.41	49.96	62.87	55.11
ECAN	MobileNet V2	53.49	43.08	35.09	45.77	45.09	44.50
CADA	MobileNet V2	62.79	53.05	43.12	49.34	59.40	53.54
SAFN	MobileNet V2	66.67	45.07	40.14	49.90	61.40	52.64
SWD	MobileNet V2	68.22	55.40	43.58	50.30	60.04	55.51
Ours	MobileNet V2	72.87	55.40	45.64	51.05	63.94	57.78

Souce Domain: AFE

Methods	Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ICID	ResNet-50	56.59	57.28	44.27	46.92	52.91	51.59
DFA	ResNet-50	51.86	52.70	38.03	41.93	60.12	48.93
LPL	ResNet-50	73.64	61.03	49.77	49.54	55.26	57.85
DETN	ResNet-50	56.27	52.11	44.72	42.17	59.80	51.01
FTDNN	ResNet-50	61.24	57.75	47.25	46.36	52.89	53.10
ECAN	ResNet-50	58.14	56.91	46.33	46.30	61.44	53.82
CADA	ResNet-50	72.09	49.77	50.92	50.32	61.70	56.96
SAFN	ResNet-50	73.64	64.79	49.08	48.89	55.69	58.42
SWD	ResNet-50	72.09	61.50	48.85	48.83	56.22	57.50
Ours	ResNet-50	78.57	65.43	51.18	51.31	62.71	61.84

Methods	Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ICID	ResNet-18	54.26	51.17	47.48	46.44	54.85	50.84
DFA	ResNet-18	35.66	45.82	34.63	36.88	62.53	43.10
LPL	ResNet-18	67.44	62.91	48.39	49.82	54.51	56.61
DETN	ResNet-18	44.19	47.23	45.46	45.39	58.41	48.14
FTDNN	ResNet-18	58.91	59.15	47.02	48.58	55.29	53.79
ECAN	ResNet-18	44.19	60.56	43.26	46.15	62.52	51.34
CADA	ResNet-18	72.09	53.99	48.39	48.61	58.50	56.32
SAFN	ResNet-18	68.22	61.50	50.46	50.07	55.17	57.08
SWD	ResNet-18	77.52	59.15	50.69	51.84	56.56	59.15
Ours	ResNet-18	79.84	61.03	51.15	51.95	65.03	61.80

Methods	Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ICID	MobileNet V2	55.04	42.72	34.86	39.94	44.34	43.38
DFA	MobileNet V2	44.19	27.70	31.88	35.95	61.55	40.25
LPL	MobileNet V2	69.77	50.23	43.35	45.57	51.63	52.11
DETN	MobileNet V2	57.36	54.46	32.80	44.11	64.36	50.62
FTDNN	MobileNet V2	65.12	46.01	46.10	46.69	53.02	51.39
ECAN	MobileNet V2	71.32	56.40	37.61	45.34	64.00	54.93
CADA	MobileNet V2	70.54	45.07	40.14	46.72	54.93	51.48
SAFN	MobileNet V2	62.79	53.99	42.66	46.61	52.65	51.74
SWD	MobileNet V2	64.34	53.52	44.72	50.24	55.85	53.73
Ours	MobileNet V2	75.19	54.46	47.25	47.88	61.10	57.18

Mean of All Methods

Souce Domain: RAF

Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ResNet-50	75.87	54.30	54.49	54.82	62.09	59.51
ResNet-18	69.43	51.88	47.94	51.72	61.26	56.45
MobileNet V2	60.78	45.15	39.59	47.92	56.46	49.98

Souce Domain: AFE

Backbone	CK+	JAFFE	SFEW2.0	FER2013	ExpW	Mean
ResNet-50	65.41	57.93	47.04	47.26	57.87	55.10
ResNet-18	60.23	56.25	46.95	47.57	58.34	53.87
MobileNet V2	63.57	48.46	40.14	44.91	56.34	50.68

Citation

@article{chen2020cross,
  title={Cross-Domain Facial Expression Recognition: A Unified Evaluation Benchmark and Adversarial Graph Learning},
  author={Chen, Tianshui and Pu, Tao and Wu, Hefeng and Xie, Yuan and Liu, Lingbo and Lin, Liang},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2021},
  pages={1-1},
  doi={10.1109/TPAMI.2021.3131222}
}

@inproceedings{xie2020adversarial,
  title={Adversarial Graph Representation Adaptation for Cross-Domain Facial Expression Recognition},
  author={Xie, Yuan and Chen, Tianshui and Pu, Tao and Wu, Hefeng and Lin, Liang},
  booktitle={Proceedings of the 28th ACM international conference on Multimedia},
  year={2020}
}

Contributors

For any questions, feel free to open an issue or contact us:

Adversarial Graph Representation Adaptation for Cross-Domain Facial Expression Recognition (AGRA, ACM 2020, Oral)

Related tags

Overview

Cross Domain Facial Expression Recognition Benchmark

Environment

Datasets

Pre-Train Model

Usage

Run ICID

Run DFA

Run LPL

Run DETN

Run FTDNN

Run ECAN

Run CADA

Run SAFN

Run SWD

Run AGRA

Result

Souce Domain: RAF

Souce Domain: AFE

Mean of All Methods

Souce Domain: RAF

Souce Domain: AFE

Citation

Contributors

Owner

CLIP2Video: Mastering Video-Text Retrieval via Image CLIP

U-Net Brain Tumor Segmentation

An Industrial Grade Federated Learning Framework

Pytorch implementation of few-shot semantic image synthesis

Tensorflow-seq2seq-tutorials - Dynamic seq2seq in TensorFlow, step by step

Tensor-based approaches for fMRI classification

Python scripts form performing stereo depth estimation using the HITNET model in Tensorflow Lite.

The challenge for Quantum Coalition Hackathon 2021

A Fast Monotone Rotating Shallow Water model

SegTransVAE: Hybrid CNN - Transformer with Regularization for medical image segmentation

Easily pull telemetry data and create beautiful visualizations for analysis.

Unofficial PyTorch implementation of Neural Additive Models (NAM) by Agarwal, et al.

LF-YOLO (Lighter and Faster YOLO) is used to detect defect of X-ray weld image.

Official PyTorch implementation of GDWCT (CVPR 2019, oral)

Implementation of association rules mining algorithms (Apriori|FPGrowth) using python.

The repository offers the official implementation of our paper in PyTorch.

Code for the paper "Attention Approximates Sparse Distributed Memory"

This repo is duplication of jwyang/faster-rcnn.pytorch

A naive ROS interface for visualDet3D.

Set of methods to ensemble boxes from different object detection models, including implementation of "Weighted boxes fusion (WBF)" method.