how can i get ROI from nvcv.ImageBatchVarShape

[lokvke]

how can i get ROI from nvcv.ImageBatchVarShape, like: x_roi = x[y1:x1, y1+height, x1 + width], or x[y1:x1, y1+height, x1 + width] = x_roi...

[bhaefnerNV]

Hi @lokvke, Thank you very much for your interest in CVCUDA!

Could you share some information on your use-case? Eg. would you like to get a ROI of the same size from an ImageBatchVarShape or would it be a different size per ImageBatchVarShape element?

[lokvke]

Hi @lokvke, Thank you very much for your interest in CVCUDA!

Could you share some information on your use-case? Eg. would you like to get a ROI of the same size from an ImageBatchVarShape or would it be a different size per ImageBatchVarShape element?

for example, i want to put the small_imgs into large_imgs' position (100, 100)

large_imgs = nvcv.ImageBatchVarShape(100)  #  shape: (100, 1280, 720, 3)
small_imgs = nvcv.ImageBatchVarShape(100)  #  shape: (100, 120, 120, 3)
x, y = (100, 100)  # (x, y)
width, height = 120, 120
large_imgs[y:y+height, x:x+width] = small_imgs

[bhaefnerNV]

Hi @lokvke,

then you could you the padandstack operator, which gets as an input an ImageBatchVarSahpe and outputs a Tensor of the cropped size. Below is a minimal working example

import cupy as cp
import numpy as np

import cvcuda

#dimensions
N = 100
height = 1280
width = 720
channels = 3
crop_height = 120
crop_width = 120
left = 100
top = 100

# create imagebatchvarshape
d_imgs = cvcuda.ImageBatchVarShape(N)
d_imgs.pushback([
    cvcuda.as_image(cp.array(np.random.rand(height, width, channels)))
    for _ in range(N)
])

# create output tensor
d_outs = cvcuda.Tensor(
    (N, crop_height, crop_width, channels),
    np.float32,
    "NHWC",
)

top_list = [-top] * N  # negative values do crop and positive do pad
left_list = [-left] * N  # negative values do crop and positive do pad

# device side
d_top = cvcuda.as_tensor(
    cp.array(top_list, dtype=np.int32).reshape(1, 1, len(top_list), 1), "NHWC")
d_left = cvcuda.as_tensor(
    cp.array(left_list, dtype=np.int32).reshape(1, 1, len(left_list), 1), "NHWC")

cvcuda.padandstack_into(dst=d_outs, src=d_imgs, top=d_top, left=d_left)

[lokvke]

Hi @lokvke,

then you could you the padandstack operator, which gets as an input an ImageBatchVarSahpe and outputs a Tensor of the cropped size. Below is a minimal working example

import cupy as cp
import numpy as np

import cvcuda

#dimensions
N = 100
height = 1280
width = 720
channels = 3
crop_height = 120
crop_width = 120
left = 100
top = 100

# create imagebatchvarshape
d_imgs = cvcuda.ImageBatchVarShape(N)
d_imgs.pushback([
    cvcuda.as_image(cp.array(np.random.rand(height, width, channels)))
    for _ in range(N)
])

# create output tensor
d_outs = cvcuda.Tensor(
    (N, crop_height, crop_width, channels),
    np.float32,
    "NHWC",
)

top_list = [-top] * N  # negative values do crop and positive do pad
left_list = [-left] * N  # negative values do crop and positive do pad

# device side
d_top = cvcuda.as_tensor(
    cp.array(top_list, dtype=np.int32).reshape(1, 1, len(top_list), 1), "NHWC")
d_left = cvcuda.as_tensor(
    cp.array(left_list, dtype=np.int32).reshape(1, 1, len(left_list), 1), "NHWC")

cvcuda.padandstack_into(dst=d_outs, src=d_imgs, top=d_top, left=d_left)

Thank you for your reply, the code is about to crop a small roi from the source images. What if i want to paste small tensors or ImageBatchVarSahpe to the source ImageBatchVarSahpe, which operator can help?