Recreating SegNet : Implementation

If you're not familiar with the SegNet algorithm, you can visit the following website.

Here, our model will contain eight levels:

• Four downscaling levels containging the following layers : Convolutional, Batch Normalisation, ReLu, Pooling.
• Four upscaling levels containing the following layers : Upsampling, Convolutional, Batch Normalisation, ReLu. We end the model by computing Softmax and a final linear layer.

Overview

• image_segmentation/: includes training and validation scripts.
• lib/: contains core functions, data preparation, model definition, and utility functions.

Installation

1. This project uses miniconda and PyTorch 1.9.1.
2. To install dependencies, run:

   pip install -r requirements.txt
   

   NOTE: TensorBorad and tensorboardX may not be compatible on some platforms. If you encounter difficulties when installing them, just remove them from requirements.txt. They are used solely for visualizing your results, which is optional (but helpful) for building your model.
3. Add current project directory (which we will later denote as ${ROOT}) to PYTHONPATH environment variable.

   export PYTHONPATH=${PYTHONPATH}:${PWD}
   

Data Preparation for Multi-digit MNIST Dataset

1. Unzip the multi-digit-MNIST dataset (multi-digit-mnist-dataset.zip) to ${ROOT}. You should have the following directory structure after unzipping:

   ${ROOT}
    `-- data
        `-- multi-digit-mnist 
            |-- batch00001.mat
            |...
            |-- testset001.mat
            |...
   
   

Training on Multi-digit MNIST Dataset

To train your model, run:

python image_segmentation/train_mnist.py

Validate the model

To validate the model after training, run:

python image_segmentation/validate_mnist.py

If you installed TensorBoard and tensorboardX, then you should have TensorBoard logs saved to out/logs. You can monitor the logs (loss curves, validation visualization, etc.) on http://localhost:6006 via:

tensorboard --logdir out/logs --port 6006

References

1. The overall structure of the code (roughly) follows Simple Baselines for Human Pose Estimation and Tracking.
2. Multi-digit MNIST dataset was created using the script from Recurrent Pixel Embedding for Instance Grouping.