Reproducing AffectNet results.

[kdhht2334]

Reproducing AffectNet results.

[kdhht2334]

1. AFEW-VA is suitable for fast check of model performance in exchange for a small number of frames. On the other hand, if you pre-process Aff-wild2's video clips, you can get about 1.4M frames. It is the most large-scale VA databse so far. Since these two datasets are from emotion recognition challenge (or variants), i think they are classified as wild DB.

   • AffectNet is also a well-refined wild DB, but the domain gap between train and test datasets is large. I think AffectNet's performance is somewhat saturated. The reason is that the SOTA 7-class accuracy (%) of this AffectNet is less than 70.

2. Many IDs can be obtained from the AffectNet dataset, but the problem is that only 1 (or 2-3) samples are given for each ID. This fact becomes a major obstacle to the implementation of ELIM, which organizes datasets by domain. Therefore, I focused on AFEW-VA and Aff-wild2 given the raw data as video clips.

• cf) So, we had no choice but to experiment with AffectNet using the age group as the domain.

[kdhht2334]

Hi,

First of all, I upload elim_category.py and update corresponding details. Please check.

Note
You must add estimated age scores to your train/val.script. And you may choose two options.

1. Use pretrained age estimation model to complete your train/val.script (link: https://github.com/yu4u/age-estimation-pytorch)
2. Borrow both age_script/training_age.csv and age_script/validation_age.csv to complete your train/val.script

• Other metadata can be obtained from AffectNet official website :)

Unfortunately, I cannot share pretrained model of AffectNet..I cannot find the weights for now.
But I can tell that you can easily get following bench-marking Table using elim_category.py!

Method	Acc.
Baseline [1]	0.58
VGG-Face [2]	0.60
HO-Conv [3]	0.59
ELIM-Age (R18)	0.611
Face2Exp [4]	0.64

References
[1] A. Mollahosseini et al., Affectnet: A database for facial expression, valence, and arousal computing in the wild, TAC 2017.
[2] D. Kollias et al., Generating faces for affect analysis, ArXiv 2018.
[3] J. Kossaifi et al., Factorized higher-order cnns with an application to spatio-temporal emotion estimation, CVPR 2020.
[4] D. Zeng et al., Face2Exp: combating data biases for facial expression recognition, CVPR 2022.

[sunggukcha]

Thanks for your reply.
My interest was Valence-Arousal prediction in AffectNet dataset.
In this case, can I get helped?

[kdhht2334]

AffectNet was mainly used for classification purposes. However, you can also use it for valence-arousal regression purposes as follows.

For convenience, I update the elim_age.py file. Please refer to this file for AffectNet.

Also i'm inspired following link:

• https://github.com/behzadhsni/BReG-NeXt

[sunggukcha]

Sincerely thanks for your kind reply.

It is rather far from the question.
"why do you prefer Aff-wild2 or AFEW-VA dataset than AffectNet (valence-arousal)?"
In my perspective of view, AffectNet has (1) more number of frames and (2) more number of identities, though it does not have temporal information.
I want your opinion on the values of the Aff-Wild and AFEW-VA.

[sunggukcha]

You helped much.
It was really nice for me to talk with you, @kdhht2334 .

Hope you happy.
Regards,
Sungguk Cha