Video Object Segmentation using Space-Time Memory Networks Seoung Wug Oh, Joon-Young Lee, Ning Xu, Seon Joo Kim [paper]

• model.py [The STM model architecture]
• video.py [Orginal forward pass function to memorize frames at an interval]

• model.py [The NEW DECODER ARCHITECTURE]
• helpers.py [helper functions for loading model parameters and segmentation scoring]
• video.py [functions to perform forward pass through video sequence]
• dataset.py [classes to construct DAVIS 2016 dataloaders]
• track.py [class to track results in the form of log file and display them]
• train.py [functions to train network in batches]
• train_DAVIS_2016.py [example script to train or validate on the DAVIS 2016 dataset.]

• python 3+
• pytorch 1.3.1+
• numpy
• tqdm
• pillow
• scikit-image

• Download the DAVIS 2016 dataset from https://davischallenge.org/davis2016/code.html
• look in train_DAVIS_2016.py for example.
• Setup the right paths to the data.
• Change settings as described.
• Run on single GPU. [tested on 1050 ti (4GB)]