Custom Crop

Overview

The Custom Crop sample demonstrates how to extract an off-center rectangular sub-region from an image using CV-CUDA’s GPU-accelerated custom crop operator. The crop region is expressed as a cvcuda.RectI (x, y, width, height) in input-image pixel coordinates, making it straightforward to implement any region-of-interest extraction pipeline.

Usage

Basic Usage

Crop the default input image with an automatically computed off-center rectangle:

python3 customcrop.py

Custom Input and Output

Specify a custom input file and output path:

python3 customcrop.py -i input.jpg -o cat_customcrop.jpg

Command-Line Arguments

Argument

Short Form

Default

Description

--input

-i

tabby_tiger_cat.jpg

Input image file path

--output

-o

cvcuda/.cache/cat_customcrop.jpg

Output image file path

Implementation

Crop Region Setup

# Determine an off-center crop rectangle.
# cvcuda.customcrop takes (x, y, width, height) in pixel coordinates,
# where (x, y) is the top-left corner of the crop region.
# We choose a region that is 60% of each dimension, starting at 20% offset
# so the crop is visually off-center and still fits within the image.
img_h, img_w = input_image.shape[0], input_image.shape[1]
crop_x = img_w // 5  # 20% from the left edge
crop_y = img_h // 5  # 20% from the top edge
crop_w = max(1, (img_w * 3) // 5)  # 60% of the image width (at least 1px)
crop_h = max(1, (img_h * 3) // 5)  # 60% of the image height (at least 1px)

# RectI specifies the crop region in the input image coordinate space
rect = cvcuda.RectI(x=crop_x, y=crop_y, width=crop_w, height=crop_h)

Applying the Custom Crop

# Determine an off-center crop rectangle.
# cvcuda.customcrop takes (x, y, width, height) in pixel coordinates,
# where (x, y) is the top-left corner of the crop region.
# We choose a region that is 60% of each dimension, starting at 20% offset
# so the crop is visually off-center and still fits within the image.
img_h, img_w = input_image.shape[0], input_image.shape[1]
crop_x = img_w // 5  # 20% from the left edge
crop_y = img_h // 5  # 20% from the top edge
crop_w = max(1, (img_w * 3) // 5)  # 60% of the image width (at least 1px)
crop_h = max(1, (img_h * 3) // 5)  # 60% of the image height (at least 1px)

# RectI specifies the crop region in the input image coordinate space
rect = cvcuda.RectI(x=crop_x, y=crop_y, width=crop_w, height=crop_h)

Key points:

  1. RectI coordinates: x and y are the top-left corner of the crop window in the input image; width and height define the output dimensions.

  2. Off-center crop: Choosing x = img_w // 5 and y = img_h // 5 deliberately avoids a centered crop, which is typical for ROI extraction use cases.

  3. Output shape: The output tensor shape is (crop_h, crop_w, channels) for an HWC input, matching exactly the rectangle dimensions.

  4. HWC layout preserved: The operator preserves the input layout (HWC or NHWC), so the result can be passed directly to downstream ops or written with write_image.

  5. Stream support: An optional stream keyword argument enables asynchronous execution on a specific CUDA stream.

Expected Output

The output shows the central 60% of the image, shifted 20% from the top-left corner:

../../_images/tabby_tiger_cat.jpg

Original Input Image
../../_images/cat_customcrop.jpg

Output: Off-center Cropped Region

CV-CUDA Operators Used

Operator

Purpose

cvcuda.customcrop()

Extract a rectangular region-of-interest from the input tensor

Common Utilities Used

See Also