SegNet is a TensorFlow implementation of the segmentation network proposed by Kendall et al., with cool features like strided deconvolution, a minified architecture and more.
Create a config.py file, containing color maps, working dataset and other options.
autoencoder = 'segnet'
colors = {
'segnet-32': [
[64, 128, 64], # Animal
[192, 0, 128], # Archway
[0, 128, 192], # Bicyclist
[0, 128, 64], # Bridge
[128, 0, 0], # Building
[64, 0, 128], # Car
[64, 0, 192], # CartLuggagePram
[192, 128, 64], # Child
[192, 192, 128], # Column_Pole
[64, 64, 128], # Fence
[128, 0, 192], # LaneMkgsDriv
[192, 0, 64], # LaneMkgsNonDriv
[128, 128, 64], # Misc_Text
[192, 0, 192], # MotorcycleScooter
[128, 64, 64], # OtherMoving
[64, 192, 128], # ParkingBlock
[64, 64, 0], # Pedestrian
[128, 64, 128], # Road
[128, 128, 192], # RoadShoulder
[0, 0, 192], # Sidewalk
[192, 128, 128], # SignSymbol
[128, 128, 128], # Sky
[64, 128, 192], # SUVPickupTruck
[0, 0, 64], # TrafficCone
[0, 64, 64], # TrafficLight
[192, 64, 128], # Train
[128, 128, 0], # Tree
[192, 128, 192], # Truck_Bus
[64, 0, 64], # Tunnel
[192, 192, 0], # VegetationMisc
[0, 0, 0], # Void
[64, 192, 0] # Wall
]
}
gpu_memory_fraction = 0.7
strided = True
working_dataset = 'segnet-32'
Two kinds of architectures are supported at the moment: the original SegNet
Encoder-Decoder (segnet), and a smaller version of the same (mini), which
can be used for simpler segmentation problems. I suggest to use strided = True
for faster and more reliable results.
The dataset_name needs to match the data directories you create in your input folder. You can use segnet-32 and segnet-13 to replicate the aforementioned Kendall et al. experiments.
Generate your TFRecords using tfrecorder.py. In order to do so, put your PNG
images in a raw folder, as follows:
input/
raw/
train/
train-labels/
test/
test-labels/
Once you have your TFRecords, train SegNet with python src/train.py. Analogously, test it with python src/test.py.