Tensorflow keras Gated-Shape CNN with MobileNetV2 backbone
Source code for Gated-Shape CNN with MobileNetV2 implementation.
Base Source code obtained from https://github.com/ben-davidson-6/Gated-SCNN
- Dataset - gated_shape_cnn/cityscape Dataset images are not uploaded to this folder as Cityscape dataset[1] is not meant for distribution.
- Dataset preparation files- gated_shape_cnn/datasets/cityscapes/
- Training scripts (loss, train scripts, final dataset preparation before training) - gated_shape_cnn/training/
Modified scripts :
- MobileNetV2 definition - gated_shape_cnn/model/mobilenetv2.py
- Shape stream layers, Fusion module layers definition - gated_shape_cnn/model/layers.py
- GSCNN model definition - gated_shape_cnn/model/model_definition.py
Jupyter notebook - GCSNN_MobileNetV2_Experiment_notebook.ipynb
Cityscapescripts[2] is installed for validation set evaluation
[1] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The Cityscapes Dataset for Semantic Urban Scene Understanding,” in Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
[2] Cityscape Evaluation scripts: https://github.com/mcordts/cityscapesScripts
Further work on model definition is required to produce good results