Project Title: Cambridge-Driving Dataset Image Segmentation using DeepLabsV3 Model with Resnet50 Backbone.
This Image segmentation project is served as an introduction to image segmentation using DeepLabsV3 through PyTorch for me. It was used to prepare for a much advanced project that can be found here. The model trained achieved a Dice score of 0.9515 after just 35 epochs. There were no subsequent post-processing steps involved in this project.
The dataset which can be found here consists of 32 classes such as cars, animal, building, etc. Pre-processing steps used in this project include;
- Normalizing: mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225] (Because of backbone)
- Resizing
- RandomHorizontalFlip (50% probability)
- RandomVerticalFlip (50% probability)
- Loss Function: Cross Enthropy Loss
- Optimizer: ADAM with learning rate of 0.001
- PyTorch
- NumPy
- PIL
- matplotlib
Extra Notes
Potential improvements could be applied to this project such as extra data augmentation, learning rate scheduling, applying post-processing e.g conditional random fields, superpixel segmentation etc. or simply letting the model training run for more epochs.