Age, gender, emotion estimation

Demo

• Instructions for demo

  python3 demo.py input_video_path="input/dinner.mp4" \

Train

• Prepare dataset
  • For age and gender: IMDB + WiKi[6]
  • For emotion: FER-2013[7]
• Instructions for train
  • If train with backone 'ShuffleNet V2', using train_shufflenet.py
  • If train with backone 'WideResNet', using train.py
  • imdb_db.mat, wiki_db.mat: please refer to [2] to run "create_db.py"
  • fer2013.csv": download from Kaggle

  python3 train_shufflenet.py
  --input_agender
  data/imdb_db.mat
  --input_wiki
  data/wiki_db.mat
  --input_emotion
  data/fer2013.csv
  --nb_epochs
  30
  --staircase_decay_at_epochs
  (5,8,)
  --lr
  0.1
  --validation_split
  0.15
  --batch_size
  64
  

References

[1] R. Rothe, R. Timofte, and L. V. Gool, "DEX: Deep EXpectation of apparent age from a single image," in Proc. of ICCV, 2015.
[2] yu4u/age-gender-estimation
[3] opconty/keras-shufflenetV2
[4] lmeulen/AgeGenderEmotion
[5] oarriaga/face_classification
[6] IMDB-WIKI – 500k+ face images with age and gender labels
[7] Challenges in Representation Learning: Facial Expression Recognition Challenge