SegNet-Basic:

What is Segnet?

• Deep Convolutional Encoder-Decoder Architecture for Semantic Pixel-wise Image Segmentation

Segnet = (Encoder + Decoder) +  Pixel-Wise Classification layer

SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation (Vijay Badrinarayanan, Alex Kendall, Roberto Cipolla, Senior Member, IEEE) arXiv:1511.00561v3

What is SegNet-Basic?

• "In order to analyse SegNet and compare its performance with FCN  (decoder  variants)  we  use  a  smaller  version  of  SegNet, termed SegNet-Basic ,  which  ha  4  encoders  and  4  decoders. All the encoders in SegNet-Basic perform max-pooling and subsampling and the corresponding decoders upsample its input using the  received   max-pooling  indices."

Basically it's a mini-segnet to experiment / test the architecure with convnets, such as FCN.

Steps To Run The Model:

1. Run python model-basic.py to create segNet_basic_model for keras to use.

• model-basic.py contains the architecure.

2. 
   

Dataset:

1. In a different directory run this to download the dataset from original Implementation.

• git clone git@github.com:alexgkendall/SegNet-Tutorial.git

• copy the /CamVid to here, or change the DataPath in data_loader.py to the above directory

2. The run python data_loader.py to generate these two files:

• /data/train_data.npz/ and /data/train_label.npz

• This will make it easy to process the model over and over, rather than waiting the data to be loaded into memory.

To Do:

[x] SegNet-Basic
[ ] SegNet
[x] Test Accuracy
[ ] Requirements

Segnet-Basic Road Scene Results:

• Train / Test:

	Train on 367 samples, validate on 233 samples
	Epoch 101/102
	366/367 [============================>.] 
	- ETA: 0s - loss: 0.3835 - acc: 0.8737Epoch 00000: val_acc improved from -inf to 0.76367, saving model to weights.best.hdf5
	367/367 [==============================] 
	- 231s - loss: 0.3832 - acc: 0.8738 - val_loss: 0.7655 - val_acc: 0.7637
	Epoch 102/102
	366/367 [============================>.] 
	- ETA: 0s - loss: 0.3589 - acc: 0.8809Epoch 00001: val_acc did not improve
	367/367 [==============================] 
	- 231s - loss: 0.3586 - acc: 0.8810 - val_loss: 2.4447 - val_acc: 0.4478

• Evaluation:

  acc: 85.47%