U-Net Implementation

By Christopher Ley

This is my interpretation and implementation of the famous paper "U-Net: Convolutional Networks for Biomedical Image Segmentation" using the Carvana Image Masking Dataset in PyTorch

This data set is a Binary Segmentation exercise of ~400 test images of cars from various angles such as those shown here:

Initial implementation for Binary Segmentation

The implementation performs almost as the winners of the competition (Dice: 0.9926 vs 0.99733) after only 5 epoch and we would expect the results to be as good as the winners using this architecture with more training and a little tweaking of the training hyper-parameters.

Here are the scores for training over 5 epochs by running:

(DeepLearning): python3 train.py

Training Results

0%|          | 0/540 [00:00<?, ?it/s]Accuracy: 103298971/467927040 = 22.08%
Dice score: 0.36127230525016785
100%|██████████| 540/540 [05:59<00:00,  1.50it/s, loss=0.0949]
==> Saving Checkpoint to: ./checkpoints/checkpoint_2022-01-06_12:39_epoch_0.pth.tar
Accuracy: 460498379/467927040 = 98.41%
Dice score: 0.9652246236801147
100%|██████████| 540/540 [05:59<00:00,  1.50it/s, loss=0.0469]
==> Saving Checkpoint to: ./checkpoints/checkpoint_2022-01-06_12:48_epoch_1.pth.tar
Accuracy: 461809183/467927040 = 98.69%
Dice score: 0.9711439609527588
100%|██████████| 540/540 [05:56<00:00,  1.51it/s, loss=0.0283]
==> Saving Checkpoint to: ./checkpoints/checkpoint_2022-01-06_12:56_epoch_2.pth.tar
Accuracy: 465675737/467927040 = 99.52%
Dice score: 0.9891990423202515
100%|██████████| 540/540 [06:00<00:00,  1.50it/s, loss=0.0194]
==> Saving Checkpoint to: ./checkpoints/checkpoint_2022-01-06_13:04_epoch_3.pth.tar
Accuracy: 465397979/467927040 = 99.46%
Dice score: 0.9878408908843994
100%|██████████| 540/540 [06:00<00:00,  1.50it/s, loss=0.0142]
==> Saving Checkpoint to: ./checkpoints/checkpoint_2022-01-06_13:12_epoch_4.pth.tar
Accuracy: 466399501/467927040 = 99.67%
Dice score: 0.9926225543022156

And an example of the output vs the ground truth of the validation set, I removed whole makes for the validation set, all 16 angles, the network had never seen this particular make from any angle.

Ground Truth

Prediction

Although limited in scope (binary segmentation for only cars), this architecture performs well with multiclass segmentation, I extended this to apply segmentation to the NYUv2 which is a multiclass objective, with little modification to the above code.

I will clean this up and upload the results and modifications soon!