Convert Kubric Optical flow Dataset

A script to convert .tfrecords of the Kubric Optical Flow Dataset to MPI Sintel Directory Structure. This directory structure makes it possible to write a PyTorch Dataloader for Kubric 'movi-f' forward and backward optical flow maps.

A PyTorch Dataloader is also provided in the EzFlow Optical Flow Library: https://github.com/neu-vi/ezflow

---

Steps

• Follow instructions on: https://github.com/google-research/kubric/tree/main/challenges/optical_flow to download the dataset.

• pip install -r requirements.txt

• python script.py --dir_path "path_where_dataset_is_saved" --save_path "path_where_converted_data_needs_to_be_saved" --split "train"

Accepted split values: train and validation

The images are saved in .png and the optical flow maps are saved in .flo format.

---

Directory Structure

If save_pathis "kubricflow"`

kubricflow
├── training
│   ├── images
│   │   ├── scene_0001
│   │   │   ├── frame_01.png
│   │   │   ├── frame_02.png
│   │   │   ├── frame_03.png
│   │   │   ├── frame_04.png
│   │   │   .
│   │   │   .
│   │   ├── scene_0002
│   │   ├── scene_0003
│   │   ├── scene_0004
│   │   .
│   │   .
│   │   .
│   ├── forward_flow
│   │   ├── scene_0001
│   │   │   ├── frame_01.flo
│   │   │   ├── frame_02.flo
│   │   │   ├── frame_03.flo
│   │   │   ├── frame_04.flo
│   │   │   .
│   │   │   .
│   │   ├── scene_0002
│   │   ├── scene_0003
│   │   ├── scene_0004
│   │   .
│   │   .
│   │   .
│   ├── backward_flow
│   │   ├── scene_0001
│   │   ├── scene_0002
│   │   ├── scene_0003
│   │   ├── scene_0004
│   │   .
│   │   .
│   │   .
├── validation
│   ├── images
│   │   ├── scene_0001
│   │   ├── scene_0002
│   │   .
│   │   .
│   ├── forward_flow
│   │   ├── scene_0001
│   │   ├── scene_0002
│   │   .
│   │   .
│   ├── backward_flow
│   │   ├── scene_0001
│   │   ├── scene_0002
│   │   .
│   │   .

---

References

• Kubric
• Kubric-github