Welcome to the public repository for our work "Learning Video Instance Segmentation with Recurrent Graph Neural Networks". The paper was published at GCPR2021 and a preliminary version is available at https://arxiv.org/pdf/2012.03911 .
We provide pre-trained weights for the backbone and detector (https://liuonline-my.sharepoint.com/:f:/g/personal/joajo88_ad_liu_se/EpVUTs9wrF1HuNiZxGyRb_kBH6lRv2YEk-7r5KFdrSFb0w?e=iY8ntF) . To run:
# Get weights
cp MY_DOWNLOADED_WEIGHTS/* pytorch_weights/.
# Download datasets.
...
# Build singularity container.
cd singularity
docker build -f Dockerfile21.09 -t pytorch21.09:211224 .
singularity build pytorch21_09.sif docker-daemon://pytorch21.09:211224
cd ..
# Setup paths
emacs singularity/rgnnvis.sh # Configure the paths
# Run training and evaluation
sbatch singularity/rgnnvis.sh runfiles/w49/rgnnvis_resnet50.py --train --ytvis_test