Low performance on AffectNet-8labels-224

[sihangchen97]

Hi,

I recently ran the test script for EmoNet on the small version dataset provided by AffectNet (291,651 images, 8 labels, cropped and resized to 224x224). However, the results I obtained were significantly lower than expected.

Expression
ACC=0.52
Valence
CCC=0.57, PCC=0.62, RMSE=0.38, SAGR=0.77
Arousal
CCC=0.52, PCC=0.56, RMSE=0.32, SAGR=0.80

I'm reaching out to confirm if this is typical performance for this version of AffectNet or if there are any adjustments I might consider to improve the results.

Should the expression_correct label also be applied to that to improve results? If so, any way I can access or generate this label for it?

Thanks for your help!