This is my implementation of Deeplab-LargeFOV.

Run the script to download the dataset:

    $ sh get_dataset.sh

Just run th training script and evaluating script:

    $ python3 train.py
    $ python3 eval.py

The super-parameters are written in the code to make it looks simple and neat. Feel free to modify them as you like if you are using this project on other datasets.

The script infer.py is an example how we could use the trained model to implement segmentation on the picture of example.jpg. Just run it to see the performance of this model:

    $ python3 infer.py

And you will see the result picture.

1. The authors claimed to use the weight pretrained on imagenet to initialize the model, and they provided the initial model which is exported from the framework of caffe. So far, I found no easy way to convert caffe pretrained weights to that of pytorch, so I initialize the model with the following methods as a remedy:

• Initialize weights of the common parts of the 'deeplab vgg' and 'pytorch standard vgg' with the weights from pytorch model zoo.

• Initialize the weights of the extraly appended layers with the msra normal distribution random numbers.

• Implement warmup with the strategy mentioned in this paper

With these methods, my naive model trained on pure pascal VOC2012 training set(1467 images) achieves the performance quite close to the performance in the paper.