If you use this code in any context, please cite the following paper:

      title={Motion Inbetweening via Deep $\Delta$-Interpolator}, 
      author={Boris N. Oreshkin and Antonios Valkanas and Félix G. Harvey and Louis-Simon Ménard and Florent Bocquelet and Mark J. Coates},

Create workspace and clone this repository

mkdir workspace

cd workspace

git clone

Build docker image and launch container

Build image and start the lightweight docker container. Note that this assumes that the data for the project will be stored in the shared folder /home/pose-estimation accessible to you and other project members.

docker build -f Dockerfile -t delta_interpolator:$USER .

nvidia-docker run -p 18888:8888 -p 16006:6006 -v ~/workspace/delta-interpolator:/workspace/delta-interpolator -t -d --shm-size="8g" --name delta_interpolator_$USER delta_interpolator:$USER

Enter docker container and launch training session

docker exec -i -t pose_estimation_$USER  /bin/bash 

Once inside docker container, this launches the training session for the proposed model. Checkpoints and tensorboard logs are stored in ./logs/lafan/transformer

python --config=src/configs/transformer.yaml

This evaluates zero-velocity and the interpolator models

python --config=src/configs/interpolator.yaml
python --config=src/configs/zerovel.yaml

Training losses eveolve as follows:

Open the results notebook to look at the metrics:



View Github