Online Meta Adaptation for Fast Video Object Segmentation

Introduction

This is the implementation of our PAMI work Online Meta Adaptation for Fast Video Object Segmentation.

Installation:

1. Clone the MVOS repository

   git clone https://github.com/huaxinxiao/MVOS-OL.git
   

2. Install - if necessary - the required dependencies:

   • Python (tested with Anaconda 2.7 and 3.6)
   • PyTorch (conda install pytorch torchvision -c pytorch - tested with PyTorch 0.3, CUDA 8.0)

Useage:

1. Download and softlink the DAVIS datasets.

   ls -s /your/davis/JPEGImages/ ./dataset/davis/
   

   For DAVIS-17, we split the multiple instances from the same video and name the file as /Annotations/480p_split.

   ls -s /your/davis/Annotations/ ./dataset/davis/
   

2. Download the pre-trained segmentation and meta models and put them under ./snapshots/.

   Pre-trained base segmentation model

   Pre-trained meta model for davis16

   Pre-trained meta model for davis17

3. Run the demo script.

   demo_mvos_davis1X.py shows the process of meta adaptation on the first frame.

   demo_mvos_ol_davis1X.py shows the process of online meta adaptation.

Results:

DAVIS16

DAVIS17

Citiation:

@article{xiao2018online,
	title={Online Meta Adaptation for Fast Video Object Segmentation},
	author={Huaxin Xiao and Bingyi Kang and Yu Liu and Maojun Zhang and Jiashi Feng},
	journal={{IEEE} Trans. Pattern Anal. Mach. Intell.},
	year={2018}
}