Training dataset contains 3936 256x256 rgb images and labels which are collected
from Capalaba, Springfield, New Farm, Fairfield, Sunnybank Hills in Brisbane, Australia.
Validation dataset contains 1344 images and labels which are collected from
Springfield and Sunnybank Hills.
Test dataset contains 1360 images and labels which are collected from a suburb in Perth, Australia.
Trained Models
SegNet 0: SegNet with 5 encoders and 5 decoders (Original SegNet).
SegNet 1: SegNet with 4 encoders and 4 decoders.
SegNet 2: SegNet with 5 encoders and 5 decoders, each encoder is replaced by a ResNet block (Block with 3 convolutional layers).
SegNet 3: SegNet with 5 encoders and 5 decoders, each encoder is replaced by a ResNet block (Block with dynamic number of convolutional layers).
Fast SCNN 0: Original Fast SCNN.
Fast SCNN 1: Fast SCNN with the first two DSConv layers removed, modified Upsample layers.
Fast SCNN 2: Fast SCNN with the first two DSConv layers replaced with Conv layers, modified Upsample layers.