lucasxlu / mmnet

Multi-task and Multi-level Feature Aggregation for Facial Expression Recognition (SOTA performance on RAF-DB)

Multi-task and Multi-level Feature Aggregation for Facial Expression Recognition

Introduction

This repository holds the PyTorch implementation of MMNet in facial expression recognition (FER) field.

Datasets

• FER2013
• RAF-DB

Data Statistics

RAF-DB Basic

Emotion	Num
0	1619
1	355
2	877
3	5957
4	2460
5	867
6	3204

FER2013

Emotion	Num
0	3995
1	436
2	4097
3	7215
4	4830
5	3171
6	4965

Performance

RAF-DB

Weighted Sampling	Race	Gender	Age	Emotion	FER(Avg Confusion Matrix)
N	0.8631	0.8217	0.7428	0.8302	0.7514
N	0.7663	0.5469	0.6281	0.8351	0.7457
Y	0.7663	0.7624	0.6004	0.8090	0.7600
Y	0.8638	0.8096	0.7396	0.8380	0.7710

FER2013

Weighted Sampling	Color	FA	Emotion Acc	Avg Confusion Matrix
N	RGB	Add	69.13	67.39
Y	RGB	Add	69.18	67.24
Y	Gray	Add	69.95	68.09
Y	Gray	Max	70.08	68.96

Ablation Study

Aggregation Mode	FER
Elw Add + Random Sampling	75.71
Elw Avg + Weighted Sampling	76.69
Elw Add + Weighted Sampling	77.10
Elw Max + Weighted Sampling	74.89
Elw Min + Weighted Sampling	75.75
None + Weighted Sampling	75.33

Learning Fashion	Race	Gender	Age	FER(Avg Confusion Matrix)
Individual Learning	86.15	82.56	73.79	73.43
Multi-task Learning	86.38	80.96	73.96	77.10

emotion_branch_w	age_branch_w	race_branch_w	gender_branch_w	Emotion_AVG_CM
1	1	1	1	0.7585
3	1	1	1	0.7710
4	1	1	1	0.7628

mmnet's Issues

pre-trained model

Hello, could you share pre-trained model in RAFDB with me? I really need it for experiments. Looking forward to your reply, thank you.