- The goal of the project is to develop a model to perform image segmentation CamVid dataset
- I will train the model on a custom dataset prepared by divamgupta. This contains video frames from a moving vehicle and is a subsample of the CamVid dataset.
- I am using a Unet neural network architecture which consists of an encoder and decoder section. This architecture is also a fully convolutional network:
- I will be using a pretrained VGG-16 network for the feature extraction path, then followed by an FCN-8 network for upsampling and generating the predictions. The output will be a label map (i.e. segmentation mask) with predictions for 12 classes.
- The model achieves an accuracy of 84% on the validation set after 170 epochs.
- I ran the notebook on Arizona State University's supercomputing cluster using one Tesla V100 GPUs. The information regarding the GPUs is included at the end of the notebook.
- Python
- Tensorflow
- Pandas
- Matplotlib
- Keras
Created by Miralireza Nabavi - feel free to contact me!