MetaVOS

Official code for the paper titled "Meta Learning Deep Visual Words for Fast Video Object Segmentation" If you use this code, please cite this Meta Learning Deep Visual Words for Fast Video Object Segmentation:

@inproceedings{DBLP:journals/corr/abs-1812-01397,
  author    = {Harkirat Singh Behl and
               Mohammad Najafi and
	       Anurag Arnab and
               Philip H. S. Torr},
  title     = {Meta Learning Deep Visual Words for Fast Video Object Segmentation},
  booktitle   = {NeurIPS 2019 Workshop on Machine Learning for Autonomous Driving},
  year      = {2018}
}

Pre Requisites

STEP-1 Installing required libraries

Pytorch version 0.4.1 is used with Python 2.7 using Anaconda2. The different libraries needed and the commands needed to install them are given in the file 'environment_setup.txt'

STEP-2 Downloading and preparing the dataset

DAVIS-17 Download the DAVIS-17 Train and Val dataset from link. After downloading the dataset, extract it within the 'metavos' directory.

For your own dataset You will need to define a new class for the dataloader for the dataset. Please refer to the class 'DavisSiameseMAMLSet' in file 'deeplab/dataset.py'.

STEP-3 Downloading the trained Model

The trained model for DAVIS-17 can be downloaded from link.

After downloading the weights, please put it in the 'snapshots' folder

TESTING

Run the file 'meta_test.py'

python meta_test.py

To visualize the segmentation, make the flag 'DISPLAY = 1' in file 'meta_test.py'

TRAINING

For Training, we start from the Pascal pretraining weights. First download the Pascal weights from link and store them in the 'snapshots' folder.

Run the file 'meta_train.py'

python meta_train.py