The pre-trained DeFMO model as reported in the paper is available here. Put the models into ./saved_models sub-folder.

For generating video temporal super-resolution:

python run.py --video example/falling_pen.avi

For generating temporal super-resolution of a single frame with the given background:

python run.py --im example/im.png --bgr example/bgr.png

Simple evaluation scripts for evaluation on FMO deblurring benchmark. You can download there all evaluation dataset: Falling Objects, TbD-3D, and TbD, which are also available here.

For the dataset generation, please download:

• ShapeNetCore.v2 dataset.

• Textures from the DTD dataset. The exact split used in DeFMO is from the "Neural Voxel Renderer: Learning an Accurate and Controllable Rendering Tool" model and can be downloaded here.

• Backgrounds for the training dataset from the VOT dataset.

• Backgrounds for the testing dataset from the Sports1M dataset.

• Blender 2.79b with Python enabled.

Then, insert your paths in renderer/settings.py file. To generate the dataset, run in renderer sub-folder:

python run_render.py

Note that the full training dataset with 50 object categories, 1000 objects per category, and 24 timestamps takes up to 1 TB of storage memory. Due to this and also the ShapeNet licence, we cannot make the pre-generated dataset public - please generate it by yourself using the steps above.

Set up all paths in main_settings.py and run

python train.py

If you use this repository, please cite the following publication:

@inproceedings{defmo,
  author = {Denys Rozumnyi and Martin R. Oswald and Vittorio Ferrari and Jiri Matas and Marc Pollefeys},
  title = {DeFMO: Deblurring and Shape Recovery of Fast Moving Objects},
  booktitle = {CVPR},
  address = {Nashville, Tennessee, USA},
  month = jun,
  year = {2021}
}