EfficientSeg: An Efficient Semantic Segmentation Network
EfficientSeg is a segmentation network using Mobilev3 blocks inside a U-shaped network structure. Using this network and the training procedure we have obtained 58.1% mIoU on Minicity test (A subset of Cityscapes) set where the baseline U-Net score was 40%.
How to run:
-
Insert Minicty train and val images under Minicity directory. To use, train+val in training and test set for testing, you can add both sets under train.
-
You can obtain the weights of the best model from : https://drive.google.com/file/d/1pZrv-LjPg3VhsU5w5MygyLHz6lCm2uvP/view
-
Using evaluate_flip.py, you can evaluate the model.
-
If you have any questions feel free to ask us via bugrayesilkaynak+efficientseg@gmail.com
-
The paper can be found here: https://arxiv.org/abs/2009.06469