Channel Reorder

Overview

The Channel Reorder sample demonstrates per-image channel permutation using CV-CUDA’s GPU-accelerated channelreorder operator. It reads an RGB image, wraps it in an ImageBatchVarShape, specifies a [2, 1, 0] channel-index order to swap R and B (producing a BGR image), and writes the result back to disk.

Usage

Basic Usage

Apply the default RGB → BGR channel swap:

python3 channelreorder.py -i input.jpg

Custom Input and Output

Specify custom input and output paths:

python3 channelreorder.py -i input.jpg -o cat_channelreorder.jpg

Command-Line Arguments

Argument

Short Form

Default

Description

--input

-i

tabby_tiger_cat.jpg

Input image file path

--output

-o

cvcuda/.cache/cat_channelreorder.jpg

Output image file path

Implementation

Setup: Wrapping the Input Tensor as an ImageBatchVarShape

# channelreorder operates on ImageBatchVarShape rather than plain tensors,
# so we wrap the HWC tensor as a cvcuda.Image and add it to a batch.
# cvcuda.as_image ties the buffer lifetime to the Image object.
cv_image: cvcuda.Image = cvcuda.as_image(
    input_image.cuda(), format=cvcuda.Format.RGB8
)

batch = cvcuda.ImageBatchVarShape(1)
batch.pushback(cv_image)

# Build the orders tensor with layout "NC" (num_images × num_channels).
# Each row lists, for every output channel, which input channel index to read.
# [2, 1, 0] maps R→B, G→G, B→R — a standard RGB-to-BGR channel swap.
order_data = np.array([[2, 1, 0]], dtype=np.int32)
orders: cvcuda.Tensor = cvcuda.Tensor((1, 3), dtype=np.int32, layout="NC")
cuda_memcpy_h2d(order_data, orders.cuda())

Channel Reorder Operator Call

# channelreorder operates on ImageBatchVarShape rather than plain tensors,
# so we wrap the HWC tensor as a cvcuda.Image and add it to a batch.
# cvcuda.as_image ties the buffer lifetime to the Image object.
cv_image: cvcuda.Image = cvcuda.as_image(
    input_image.cuda(), format=cvcuda.Format.RGB8
)

batch = cvcuda.ImageBatchVarShape(1)
batch.pushback(cv_image)

# Build the orders tensor with layout "NC" (num_images × num_channels).
# Each row lists, for every output channel, which input channel index to read.
# [2, 1, 0] maps R→B, G→G, B→R — a standard RGB-to-BGR channel swap.
order_data = np.array([[2, 1, 0]], dtype=np.int32)
orders: cvcuda.Tensor = cvcuda.Tensor((1, 3), dtype=np.int32, layout="NC")
cuda_memcpy_h2d(order_data, orders.cuda())

Key points:

  1. Two input containers — fixed-shape Tensor inputs take a host order sequence, while ImageBatchVarShape inputs take a device-resident per-image orders tensor.

  2. VarShape orders layout — the orders tensor must have layout "NC" (num_images × num_channels), where each row gives the input-channel index for each output channel.

  3. Tensor order sequence — cvcuda.channelreorder(tensor, [2, 1, 0]) performs the canonical RGB → BGR swap without allocating a device parameter tensor.

  4. Zero-copy wrapping — cvcuda.as_image ties the source buffer lifetime to the Image object; no extra device copy is performed.

  5. Result extraction — iterate over a variable-shape output and call cvcuda.as_tensor(out_image.cuda(), "HWC") to obtain a writable HWC tensor.

Expected Output

The output image has its red and blue channels swapped relative to the input:

../../_images/tabby_tiger_cat.jpg

Original Input Image (RGB)
../../_images/cat_channelreorder.jpg

Output: Channels Reordered to BGR

CV-CUDA Operators Used

Operator

Purpose

cvcuda.channelreorder()

Permute image channels according to a per-image index tensor

Common Utilities Used

  • read_image() - Load image as CV-CUDA tensor

  • write_image() - Save reordered image

  • cuda_memcpy_h2d - Upload the host-side orders array to the GPU

See Also