Natural movies

Description of the data and file structure

Various types of natural scenes were obtained, categorized as flow (camera moving through space), pan (camera panning through space), still (fixed camera recording a still scene) and still moving (fixed camera recording an active scene). A total of 39 clips of 9.6 second duration were recorded at a frame rate of 120 FPS and a spatial extent of 720 × 1280. The dataset is divided into a training and validation set, both containing a balanced number of the different recording types (training set: 12 flow, 8 pan, 3 still and 8 still moving clips; validation set: 3 flow, 2 pan, 1 still and 2 still moving clips). All clips were preprocessed by converting the RGB channels to grayscale and bilinearly downsampling the spatial dimensions to 144 × 256.

Code/Software

Data is stored h5py format and can be accessed in Python as follows:

import h5py

root = "" # path to data
hf = h5py.File(f"{root}/natural.hdf5", "r")
dataset_name = "train" # or "test"
dataset = np.array(hf.get(dataset_name))
hf.close()