[DIP] Make Resize API more consistent with OpenCV#154
Conversation
|
Sorry, i misunderstood OpenCV doc, its resize() read params in [width, height] format, instead of [h, w]. I 'm really sorry for this mistake, later i will revise the initial func in OpenCV benchmark design. |
|
Shall i delete |
| "input_image_dimension / output_image_dimension\n"); | ||
| "output_image_dimension / input_image_dimension\n"); | ||
| } | ||
| std::reverse(scalingRatios.begin(), scalingRatios.end()); |
There was a problem hiding this comment.
Also adjust the OpenCV dimension(previous PR give [h, w], but it need [w, h]), so the dimension conversion is no longer required.
There was a problem hiding this comment.
I am not sure I understand, we need to feed dimensions in [h, w] format at the lower level and we capture input in [w, h] format from the higher level.
There was a problem hiding this comment.
The benchmark code acquire inputs in [h, w], in the example <RowNum> comes before <ColNum>, so we capture high level inputs in [h, w] format:
./image-processing-resize-benchmark <image path> <Scale option> <RowNum> <ColNum> <InterpolationOption>
The lower level format is only determined by API(Buddy or OpenCV), so both of them need to convert [h, w] input in high level to [w, h] level in low level.
How about changing the benchmark code, i will do the [h, w] -> [w, h] conversion in the benchmark code, so the low level API get the [w, h] format. Also, the code in DIP.h dont need a switch anymore, the following code
return detail::Resize2D_Impl( input, type, {scalingRatios[1], scalingRatios[0]}, outputSize);
can change to
return detail::Resize2D_Impl( input, type, {scalingRatios[0], scalingRatios[1]}, outputSize); ?
There was a problem hiding this comment.
Please change the benchmark code to take input in the [w, h] format.
Applying std::reverse makes it easier to read the code imo, because then we use scalingRatio[0] to get outputSize[0] and scalingRatio[1] to get outputSize[1].
There was a problem hiding this comment.
Please change the benchmark code to take input in the [w, h] format.
Please have a review at the resize2d PR. The tasks mentioned above were in the new commits.
So does this proposed API :)
Okay |
2 participants
This PR intends to make some changes in resize API to make it more consistent with OpenCV. Ref: buddy-compiler/buddy-benchmark#61 (comment)
@taiqzheng, can you confirm if these changes would help your work?