Fast library to create toy optical flow datasets on fly to test your deep learning models.

Why?

If you want to sanity test 🧪 your deep learning model & don't want to spend time ⌚ & effort 🏋️‍♀️ to run one full iteration on the "Flying Chairs" dataset, you can use this library to generate (easier, smaller & customizable) toy optical flow datasets.

What is an optical flow dataset ❓

Each data point in an optical flow dataset (like FLying Chairs) consist of 3 things. Image-1 📷, Image-2 📷 & an array of shape (Height x Width x 2) which stores the optical flow b/w Image-1 & Image-2
Standard optical flow datasets are big & harder 🔴 to test with (& rightly so, since these datasets are based on real or close to real life images & scenarios).

To any one wondering what "flying chair" is? It is a standard dataset that is used to compare performance of optical flow models (https://lmb.informatik.uni-freiburg.de/resources/datasets/FlyingChairs.en.html#flyingchairs). Think of it as being the GLUE/CIFAR‑100 of optical flow estimation research.

Code overview 👩‍💻

data_shapes.py is used to create datapoints where a single shape is moving.
data_shapes_double.py is used to create datapoints with 2 shapes, one can customize the % of occlusion to vary the difficulty of the points.

Each datapoint consist of two images with a randomly generated shape imposed on a black background and its calculated optical flow.