flow_utils.py

Basic functions to handle optical flow.

ptlflow.utils.flow_utils.flow_read(input_file: str | Path | IO, format: str | None = None) → ndarray[source]

Read optical flow from file.

This is just a wrapper for flowpy (for .flo and .png) or raft (for pfm), added for convenience.

Parameters:

input_file: str, pathlib.Path or IO: Path of the file to read or file object.
format: str, optional: Specify in what format the flow is read, accepted formats: “png”, “flo”, “pfm”, or “flo5”. If None, it is guessed on the file extension.

Returns:

numpy.ndarray: 3D flow in the HWF (Height, Width, Flow) layout. flow[…, 0] is the x-displacement. flow[…, 1] is the y-displacement.

See also

Convert flows to RGB images.

The input can be either numpy or torch tensors. This function is just a wrapper for flowpy and flowpy_torch.

Parameters:

flownp.ndarray or torch.Tensor: If flow is a numpy array, then it must have 3 dimensions HWC (Height, Width, Channels) - notice it is channels last. If it is a torch tensor, then it has at least 3 dimensions in the …CHW (…, Channels, Height, Width) layout, where … represents any number of dimensions. Channel 0 should be the x-displacement. Channel 1 should be the y-displacement.
flow_max_radiusfloat or torch.Tensor, optional: Set the radius that gives the maximum color intensity, useful for comparing different flows. Default: The normalization is based on the input flow maximum radius per batch element.
backgroundstr, default ‘bright’: States if zero-valued flow should look ‘bright’ or ‘dark’.
custom_colorwheelnp.ndarray or torch.Tensor: Use a custom colorwheel for specific hue transition lengths. By default, the default transition lengths are used.

Returns:

np.ndarray or torch.Tensor: The RGB image representing the flow. It keeps the same dimensions and type as the input.