Why saved out (to Mat) interpolated tensor is a grid image

vision
1

Below is my code.

Mat img = imread(“original.jpg”);
auto img_tensor = torch::from_blob(img.data, {img.rows, img.cols, img.channels()}, at::kByte);
//Save out orignial tensor to check
Mat saveOutOriginal(img_tensor.sizes()[0], img_tensor.sizes()[1], CV_8UC3);
memcpy((void*)saveOutOriginal.data, img_tensor.data_ptr(), sizeof(torch::kU8)img_tensor.numel());
imwrite(“original_copy.jpg”, saveOutOriginal);
//Reshape to do interpolation
img_tensor = img_tensor.permute({2,0,1});
img_tensor = img_tensor.unsqueeze(0);
img_tensor = img_tensor.to(torch::kFloat);
auto resized_tensor = functional::interpolate(img_tensor, functional::InterpolateFuncOptions().size(vector<int64_t>({512,512})).mode(torch::kBilinear));
resized_tensor = resized_tensor.squeeze().detach().permute({1,2,0});
resized_tensor = resized_tensor.to(torch::kU8);
//Save out resized result to check
Mat checkResized(resized_tensor.sizes()[0], resized_tensor.sizes()[0], CV_8UC3);
memcpy((void)checkResized.data, resized_tensor.data_ptr(), sizeof(torch::kU8)*resized_tensor.numel());
imwrite(“resized.jpg”, checkResized);

I check the unchanged tensor “original.jpg”. It’s the same with input image.
The result is as expected.
But after resizing the tensor and save it out as a Mat.
Why it becomes 9 times grid image (see below) ?
How can I get output with only one resized image (512x512) ?

original_copy.jpg
original_copy

resized.jpg

resized
resized512×512 65.2 KB

2

I dont know why, but I solved my problem by spliting the saved mat and merging .

int width = 512;
int height = 512;
uchar* data_encode = checkResized.data;
cv::Mat b(height,width,CV_8UC1,&data_encode[0]);
cv::Mat g(height,width,CV_8UC1,&data_encode[(heightwidth)sizeof(uchar)]);
cv::Mat r(height,width,CV_8UC1,&data_encode[(2heightwidth)*sizeof(uchar)]);
test.push_back(b);
test.push_back(g);
test.push_back( r );
cv::Mat out ;
cv::merge(test,out);

The final output out is correct.
For your reference.