Skip to content

Latest commit

 

History

History
40 lines (25 loc) · 1.32 KB

README.md

File metadata and controls

40 lines (25 loc) · 1.32 KB
Raw

Facial Expression Recognition

Datasets

AffectNet

  • AffectNet-7: 283,901 train set, 3,500 test set and 7 labels
  • AffectNet-8: 287,568 train set, 4,000 test set and 8 labels

RAF-DB

  • 12,271 train set and 3,068 test set
  • 7 baisc and 11 compound emotion labels
  • Faces are aligned and cropped to 100x100 size.

SFEW 2.0

  • Newest version of SFEW dataset
  • Extracted from AFEW video database
  • 7 expression categories
  • 958 train set, 436 val set and 372 test set
  • Light and compact
  • Faces are not aligned.

Pre-training

Most facial expression recognition models are pre-trained on MS-Celeb-1M dataset. Download the pretrained model from here.

Benchmarks & Pretrained Models

Model Backbone AffectNet-8 AffectNet-7 RAF-DB SFEW 2.0 Params (M) GLOFPs Checkpoints
DAN ResNet-18 62.09 65.69 89.70 57.88 - - AffectNet8 | AffectNet7 | RAFDB