FAMINet

Implementation of our paper entitiled FAMINet: Learning Real-time Semi-supervised Video Object Segmentation with Steepest Optimized Optical Flow published in TIM. This code is mainly based on frtm-vos. Thanks for their provided codes.

1. Get Started

a. Test environment:

Ubuntu 16.04
Python 3.7
Pytorch 1.7

b. Install:

sudo apt install ninja-build
pip install scipy scikit-image tqdm opencv-python easydict

2. Datasets

a. DAVIS dataset

DAVIS dataset is from the DAVIS benchmark: https://davischallenge.org/davis2017/code.html. Users can directly download DAVIS2017 from https://data.vision.ee.ethz.ch/csergi/share/davis/DAVIS-2017-trainval-480p.zip.

b. Youtube-VOS dataset

Youtube-VOS dataset is from the Youtube-VOS benchmark: https://youtube-vos.org/dataset/.

c. Set path

After downloading the overall datasets, set the corresponding dataset path in evaluate.py and train.py, e.g.:

davis="./DAVIS",  # DAVIS dataset root
yt="/data3/YouTubeVOS",  # YouTubeVOS root
output="./results", # output results

3. Test

python evaluate.py --model model_path --fast --dset ytval # YouTubeVos
python evaluate.py --model model_path --fast --dset dv2016val   # DAVIS 2016
python evaluate.py --model model_path --fast --dset dv2017val   # DAVIS 2017

We provided our model FAMINet-2F and FAMINet-3F for reference:

Name	Backbone	Weights
FAMINet-2F, FAMINet-3F	ResNet18	Download

4. Train

python train.py name --ftext resnet18 --dset all --dev gpu_id

name experiment name.

dset dataset used for training, e.g., DAVIS, Youtube2018.