U2-Net

A Pytorch implementation of U2-Net: Going Deeper with Nested U−Structure for Salient Object Detection trained with CelebAMask

1. Train with Matting Human Dataset (2022.08.01)

I made a dataset and upload in Kaggle after pre-processing. (288x384 & 216x288)

Implementation Details

augmentation: 288x384 → RandomHorizontalFlip → RandomCrop → 256x256
optimizer: Adam (learning rate = 1e-4) 
epoch: 25
loss weight: all 1
batch size: 12

train loss: binary cross entrophy
validation & test loss: mean absolute error (L1 loss)

Output of test dataset images

Convert GIFs

Loss

Train loss

Valid loss

- minimum: 0.0077

Test loss

Mean absolute error for test images: 0.00806

2. Train with CelebAMask (2022.06.24)

I made a dataset and upload in Kaggle after pre-processing. (128x128 & 256x256)

Implementation Details

image size: 128x128
optimizer: Adam (learning rate = 1e-3)
epoch: 25
loss weight: all 1
batch size: 12

train loss: binary cross entrophy
validation & test loss: mean absolute error (L1 loss)

Output of test dataset images

Convert GIFs

Loss

Train loss

Test loss

Mean absolute error for test images: 0.02649