Color Conversion

Overview

The Color Conversion sample demonstrates GPU-accelerated color space conversion using CV-CUDA’s cvtcolor operator. The example reads an input image, converts it from RGB to BGR by swapping the red and blue channels, and writes the result as a viewable uint8 JPEG. The same operator supports a wide range of conversions including grayscale, RGBA, HSV, and YUV formats — only the code argument needs to change.

Usage

Basic Usage

Convert the default tabby-cat image (RGB to BGR):

python3 cvtcolor.py

Custom Input and Output

Specify input and output paths explicitly:

python3 cvtcolor.py -i input.jpg -o cat_cvtcolor.jpg

Command-Line Arguments

Argument

Short Form

Default

Description

--input

-i

tabby_tiger_cat.jpg

Input image file path

--output

-o

cvcuda/.cache/cat_cvtcolor.jpg

Output image file path

Implementation

Color Space Conversion

# cvtcolor requires a batched (NHWC) tensor, so wrap the HWC image in a batch
# dimension using cvcuda.stack before passing it to the operator.
nhwc_image: cvcuda.Tensor = cvcuda.stack([input_image])

# Swap the R and B channels (RGB2BGR) — the output is a visually distinct
# but still fully viewable 3-channel uint8 image, making the conversion easy
# to verify by eye (warm tones shift to cool and vice versa).
converted: cvcuda.Tensor = cvcuda.cvtcolor(
    nhwc_image, code=cvcuda.ColorConversion.RGB2BGR
)

# Drop the batch dimension back to HWC so write_image can encode the result.
output_image: cvcuda.Tensor = converted.reshape(converted.shape[1:], "HWC")
write_image(output_image, args.output)

Key points:

  1. Batched input: cvcuda.cvtcolor requires an NHWC tensor; a single HWC image is promoted to a batch of one with cvcuda.stack.

  2. ColorConversion enum: The desired conversion is selected by passing a :pydata:`cvcuda.ColorConversion` member as the code keyword argument.

  3. Symmetric channel counts: The source and destination channel counts must match the chosen conversion code (e.g. RGB2BGR keeps 3 channels; BGR2GRAY reduces to 1).

  4. Batch dimension removal: After conversion the leading batch dimension is dropped with Tensor.reshape so the result is a plain HWC tensor that write_image can encode directly as JPEG.

  5. Supported dtypes: The operator accepts uint8 and uint16 inputs; the default JPEG pipeline uses uint8.

Expected Output

The output shows the input image with red and blue channels exchanged. Warm-toned areas (e.g. orange fur) appear cooler and vice versa:

../../_images/tabby_tiger_cat.jpg

Original Input Image
../../_images/cat_cvtcolor.jpg

Output: RGB channels converted to BGR

CV-CUDA Operators Used

Operator

Purpose

cvcuda.cvtcolor()

Convert image between color spaces using a GPU-accelerated kernel

Common Utilities Used

See Also