Converted MagFace / AdaFace r50 / r100 model and face quality testing

[leondgarse]

Magface

• Model weights ported from Github IrvingMeng/MagFace. Paper MagFace: A Universal Representation for Face Recognition and Quality Assessment. These models can be used testing face quality.

• Model input is BGR image, value in [0, 1].

• Model architecture is:

  import models
  mm = models.buildin_models('r50', output_layer='E', activation="PReLU", bn_momentum=0.9, bn_epsilon=2e-5, use_bias=True, scale=True)

• Evaluation on bin test files:

  import evals
  bb = keras.models.load_model('r50_magface_MS1MV2.h5')
  
  eea = evals.eval_callback(lambda imms: bb(imms[:, :, :, ::-1] / 2 + 0.5), '/datasets/ms1m-retinaface-t1/cfp_fp.bin', batch_size=16)
  eea.on_epoch_end()
  # >>>> cfp_fp evaluation max accuracy: 0.981143, thresh: 0.196732
  
  eeb = evals.eval_callback(lambda imms: bb(imms[:, :, :, ::-1] / 2 + 0.5), '/datasets/ms1m-retinaface-t1/agedb_30.bin', batch_size=16)
  eeb.on_epoch_end()
  # >>>> agedb_30 evaluation max accuracy: 0.980500, thresh: 0.255450
  
  """ Plot face quality distribution using norm value of feature """
  cc = tf.norm(eea.embs, axis=1).numpy()
  _ = plt.hist(cc, bins=512, alpha=0.5, label='cfp_fp quality')
  dd = tf.norm(eeb.embs, axis=1).numpy()
  _ = plt.hist(dd, bins=512, alpha=0.5, label='agedb_30 quality')
  plt.legend()
  plt.tight_layout()

  

• Evaluation on IJBB

  import IJB_evals
  mm = keras.models.load_model('r50_magface_MS1MV2.h5')
  tt = IJB_evals.IJB_test(lambda imms: mm(imms[:, :, :, ::-1] / 255.0), data_path='/datasets/IJB_release/', subset='IJBB', batch_size=16)
  score = tt.run_model_test_single()
  IJB_evals.plot_roc_and_calculate_tpr([score], names=[mm.name + "_IJBB"], label=tt.label)

• Results:

  Model name in IrvingMeng/MagFace	lfw	cfp_fp	agedb_30	IJBB
  magface_iresnet50_MS1MV2_ddp_fp32	0.997500	0.981143	0.978833	0.930185
  mag-cosface_iresnet50_MS1MV2_ddp_fp32	0.998333	0.982714	0.978667	0.93408
  magface_iresnet50_MS1MV2_dp	0.998167	0.981143	0.980500	0.943622
  magface_epoch_00025	0.998333	0.987429	0.983333	0.949562

• Shared r50_magface_MS1MV2.h5 is magface_iresnet50_MS1MV2_dp, r100_magface_MS1MV2.h5 is magface_epoch_00025.

[leondgarse]

Adaface

• Model weights ported from Github mk-minchul/AdaFace. Paper AdaFace: Quality Adaptive Margin for Face Recognition. These models can be used testing face quality. Original model input format is BGR image value in [-1, 1]. Model first Conv layer weights are converted fitting RGB input.

• Model architecture is:

  import models
  mm = models.buildin_models('r100', output_layer='E', activation="PReLU", bn_momentum=0.9, bn_epsilon=1e-5, use_bias=True, scale=False, use_max_pool=True)

• Evaluation on bin test files:

  import evals
  mm = keras.models.load_model('adaface_ir101_webface12m_rgb.h5')
  
  eea = evals.eval_callback(mm, '/datasets/ms1m-retinaface-t1/cfp_fp.bin', batch_size=16)
  eea.on_epoch_end()
  # >>>> cfp_fp evaluation max accuracy: 0.993286, thresh: 0.181749
  
  eeb = evals.eval_callback(mm, '/datasets/ms1m-retinaface-t1/agedb_30.bin', batch_size=16)
  eeb.on_epoch_end()
  # >>>> agedb_30 evaluation max accuracy: 0.981667, thresh: 0.194335
  
  """ Plot face quality distribution using norm value of feature """
  cc = tf.norm(eea.embs, axis=1).numpy()
  _ = plt.hist(cc, bins=512, alpha=0.5, label='cfp_fp quality')
  dd = tf.norm(eeb.embs, axis=1).numpy()
  _ = plt.hist(dd, bins=512, alpha=0.5, label='agedb_30 quality')
  plt.legend()
  plt.tight_layout()

  

• Evaluation

  CUDA_VISIBLE_DEVICES='0' ./IJB_evals.py -m adaface_ir101_webface12m_rgb.h5 -d /datasets/IJB_release/

  Model	lfw	cfp_fp	agedb_30	IJBB	IJBC
  adaface_ir101_webface4m	0.998333	0.992857	0.978833	0.960954	0.974485
  adaface_ir101_webface12m	0.998500	0.993286	0.981667	0.964752	0.977451

  IJBB / IJBC detail

  Model	1e-06	1e-05	0.0001	0.001	0.01	0.1	AUC
  adaface_ir101_webface4m, IJBB	0.451412	0.929309	0.960954	0.97371	0.983057	0.99036	0.995527
  adaface_ir101_webface4m, IJBC	0.922585	0.961395	0.974485	0.983177	0.988853	0.993404	0.996997
  adaface_ir101_webface12m, IJBB	0.490068	0.936027	0.964752	0.974586	0.982571	0.990555	0.995308
  adaface_ir101_webface12m, IJBC	0.900854	0.963952	0.977451	0.983944	0.988597	0.993148	0.996642