Paper: U-Net: Convolutional Networks for Biomedical Image Segmentation
Dataset used: Semantic segmentation drone dataset
This implementation follows the same architecture as in the paper, only the last 3x3 convolution is removed (can be seen in unet.py).
Output has 5 channels of logits which will give probabilities upon performing softmax on them.
The below example is from a model trained on 300 of the images from the dataset, with epochs = 205, batch size = 8, and learning rate = 0.0005.
| Original Image | Predicted output | Expected output |
|---|---|---|
 |
 Note: The prediction is a center crop. |
 |
This model's parameters are available here (0.0271 in the name is the cost of the model on the training set).