Convert To

Overview

The Convert To sample demonstrates dtype conversion using CV-CUDA’s GPU-accelerated convertto operator. It converts a uint8 image to float32 with a scale factor of 1/255 (normalising pixel values to [0, 1]), then converts the result back to uint8 by applying the inverse scale of 255. This round-trip is a fundamental pre/post-processing step for deep-learning inference pipelines.

Usage

Basic Usage

Convert the default tabby cat image:

python3 convertto.py

Custom Input/Output

Specify explicit input and output paths:

python3 convertto.py -i image.jpg -o cat_convertto.jpg

Command-Line Arguments

Argument	Short Form	Default	Description
`--input`	`-i`	tabby_tiger_cat.jpg	Input image file path
`--output`	`-o`	cvcuda/.cache/cat_convertto.jpg	Output image file path

Implementation

Convert To Operator

# 1. Convert the uint8 image to float32, applying a scale factor.
#    scale=1/255.0 maps [0, 255] -> [0.0, 1.0] — a standard normalization step
#    used before feeding images into neural networks.
float_image: cvcuda.Tensor = cvcuda.convertto(
    src=input_image,
    dtype=np.float32,
    scale=1.0 / 255.0,
)

# 2. Convert back to uint8 by reversing the scale (multiply by 255).
#    This round-trip demonstrates that the conversion is lossless for
#    images with pixel values in the valid uint8 range.
output_image: cvcuda.Tensor = cvcuda.convertto(
    src=float_image,
    dtype=np.uint8,
    scale=255.0,
)

Key points:

dtype parameter: Pass a numpy dtype (e.g. np.float32) or a cvcuda.Type enum value — both are accepted by cvcuda.convertto.
scale parameter: Each output pixel is computed as out = src * scale + offset. Omitting scale defaults to 1.0.
offset parameter: An optional additive bias applied after scaling; defaults to 0.0 when omitted.
Layout preservation: The output tensor always has the same layout (HWC, NHWC, CHW, NCHW) as the input tensor.

Expected Output

The output image is visually identical to the input because the uint8→float32→uint8 round-trip is lossless for pixel values in [0, 255]:

../../_images/tabby_tiger_cat.jpg — Original Input Image

../../_images/cat_convertto.jpg — Output: uint8 round-trip via float32

CV-CUDA Operators Used

Operator	Purpose
`cvcuda.convertto()`	Convert tensor dtype with optional scale and offset

Common Utilities Used

read_image() - Load image as CV-CUDA tensor
write_image() - Save converted image