Skip to content

Augmentation effect

Viet edited this page Jun 12, 2019 · 2 revisions

Baseline Model

With Histogram Equalization

  • General parameter:
Parameter Metrics
batch_size: 32 f1-score: -
hist_equalization: True mse: -
image_resolution:(512, 512) optimal mse threshold: -
num_epochs: 1000 roc-auc: 0.4541

Without Histogram Equalization

  • General parameter:
Parameter Metrics
batch_size: 32 f1-score: -
hist_equalization: True mse: -
image_resolution:(512, 512) optimal mse threshold: -
num_epochs: 1000 roc-auc: 0.4541

Augmentation Effects

Augmentation 1: 5 different methods

  • General parameter:
Parameter Metrics
batch_size: 32 f1-score: 0.737247353224254
hist_equalization: True mse: 0.0024145432
image_resolution:(512, 512) optimal mse threshold: 0.00030937084
num_epochs: 1000 roc-auc: 0.4567827733624078
  • Augmentation parameter:
Chance Parameter Augmentation
1 iaa.Fliplr(0.5) Flip horizontally with 50%
1 iaa.Flipud(0.1) Flip vertically with 10%
0.5 iaa.Affine(rotate=(-15, 15)) Rotate from -15° to 15°
0.5 iaa.Multiply((0.8, 1.2)) Multiply Values from 0.8 to 1.2 → augment brighness
0.5 iaa.Affine(scale=(0.8, 1.2))) Zoom in and out from 0.8 to 1.2
  • Last image: augmentation_1_last_image

  • Loss

    • Does not overfit.
    • Some peaks in the training loss
    • Validation loss is twice as high as training loss → train longer

Loss_Augmentation_1

  • False positive: highest MSE
    • Have a lot of noisy space around the hand
    • Not common angles
    • Black image

false_positive_augmentation_1

  • False negative: lowest MSE
    • Common angles
    • Space around the hand is not to noisy

false_negative_augmentation_1

  • True negative: lowest MSE

true_negative_augmentation_1

  • True positive: highest MSE
    • Images with a lot of noise
    • Uncommon angles

true_positive_augmentation_1

Augmentation 2: Flip Hand

  • General parameter:
Parameter Metrics
batch_size: 32 f1-score: 0.7373493975903614
hist_equalization: True mse: 0.0021393178
image_resolution:(256, 256) optimal mse threshold: 0.0005970913
num_epochs: 500 roc-auc: 0.45168714041362146
  • Augmentation parameter:
Parameter Augmentation
iaa.Fliplr(0.5) Flip horizontally with 50%
  • Last image

augmentation_2_last_image

  • Loss
    • Does not overfit.

Loss_Augmentation_2

  • False positive: highest MSE
    • Have a lot of noisy space around the hand
    • Not common angles
    • Black image

false_positive_augmentation_2

  • False negative: lowest MSE
    • Common angles
    • Space around the hand is not to noisy

false_negative_augmentation_2

  • True negative: lowest MSE

true_negative_augmentation_2

  • True positive: highest MSE
    • Images with a lot of noise
    • Uncommon angles

true_positive_augmentation_2