mpcompress.utils.transforms

Color space conversion utilities. Adapted from https://github.com/microsoft/DCVC/blob/main/src/utils/transforms.py

rgb2ycbcr

rgb2ycbcr(rgb: Tensor, is_bgr=False)

Convert RGB tensor to YCbCr color space.

Converts RGB values to YCbCr using ITU-R BT.709 standard weights. The output values are clamped to [0, 1] range.

Parameters:

Name	Type	Description	Default
`rgb`	`Tensor`	RGB tensor with shape (..., 3, H, W) or (..., C, H, W) where the last 3 channels are RGB.	required
`is_bgr`	`bool`	If True, interpret input as BGR instead of RGB. Defaults to False.	`False`

Returns:

Name	Type	Description
`ycbcr`	`Tensor`	YCbCr tensor with shape (..., 3, H, W), values clamped to [0, 1].

ycbcr2rgb

ycbcr2rgb(ycbcr: Tensor, is_bgr=False, clamp=True)

Convert YCbCr tensor to RGB color space.

Converts YCbCr values to RGB using ITU-R BT.709 standard weights. Optionally clamps output values to [0, 1] range.

Parameters:

Name	Type	Description	Default
`ycbcr`	`Tensor`	YCbCr tensor with shape (..., 3, H, W) or (..., C, H, W) where the last 3 channels are YCbCr.	required
`is_bgr`	`bool`	If True, output as BGR instead of RGB. Defaults to False.	`False`
`clamp`	`bool`	If True, clamp output values to [0, 1] range. Defaults to True.	`True`

Returns:

Name	Type	Description
`rgb`	`Tensor`	RGB tensor with shape (..., 3, H, W), optionally clamped to [0, 1].

ycbcr420_to_444_np

ycbcr420_to_444_np(y, uv, order=0, separate=False)

Convert YCbCr 4:2:0 format to 4:4:4 format using numpy.

Upsamples the chroma channels (UV) from half resolution to full resolution to match the luma channel (Y) resolution.

Parameters:

Name	Type	Description	Default
`y`	`ndarray`	Y channel float numpy array with shape (1, H, W).	required
`uv`	`ndarray`	UV channels float numpy array with shape (2, H/2, W/2).	required
`order`	`int`	Interpolation order for upsampling. 0 for nearest neighbor (default), 1 for bilinear interpolation.	`0`
`separate`	`bool`	If True, return Y and UV separately. If False, return concatenated YCbCr array. Defaults to False.	`False`

Returns:

Name	Type	Description
`out`	`ndarray or tuple(ndarray, ndarray)`	If separate is False, return yuv with shape (3, H, W). If separate is True, return (y, uv) with shapes (1, H, W) and (2, H/2, W/2).

yuv_444_to_420

yuv_444_to_420(yuv)

Convert YUV 4:4:4 format to 4:2:0 format.

Downsamples the chroma channels (UV) from full resolution to half resolution using average pooling, while keeping the luma channel (Y) at full resolution.

Parameters:

Name	Type	Description	Default
`yuv`	`Tensor`	YUV tensor with shape (B, 3, H, W) where channels are (Y, Cb, Cr).	required

Returns:

Name	Type	Description
`y`	`Tensor`	Y channel with shape (B, 1, H, W).
`uv`	`Tensor`	Downsampled UV channels with shape (B, 2, H/2, W/2).