This demo shows how to train and run a Semantic Segmentation network for the "Person vs. Background" task. The model is trained on a subset of the COCO segmentation dataset (http://cocodataset.org/), containing "person" class. The network architecture is made from scratch and is inspired be U-Net and DeepLab v3 (ASPP) architectures
person-segmentation.ipynb- All in one notebook: load data, create model, run training, run inference
post-processing.ipynb- Playground for segmentation post processing (applications)
- Emulating DOF effect and background switch
kaggle_submission.ipynb- Run all test images through a model
- Generate a kaggle submission file (convers all segmentation maps to CSV using RLE)
The scripts contain approximatelly the same code, as in person-segmentation.ipynb
data.py- Dataset classes for preparation fo training and validation data pipelines
model.py- Definition of the Model class, defining the network architecture
train.py- Execute this script to run the training procedure
inference.py- Execute this script to run the inference on a trained model