NVIDIA / NVIDIA/TensorRT-Edge-LLM
stbir_resize in resizeImage maybe calc wrong result
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 563
- Forks
- 135
- Avg merge
- 14h 13m
- Merged PRs (30d)
- 1
Description
stbir_resize in resizeImage maybe calc wrong result
if (doResize)
{
auto [resizedHeight, resizedWidth] = getResizedImageSize(image.height, image.width);
rt::imageUtils::resizeImage(
image, mResizedImageHost, resizedWidth, resizedHeight, rt::imageUtils::InterpolationMode::kBICUBIC);
formatPatch(mResizedImageHost, imageGridTHWs, imageTokenLengths, cuSeqlensData, cuSeqlensSize,
maxSeqLen, stream);
}
else
{
formatPatch(image, imageGridTHWs, imageTokenLengths, cuSeqlensData, cuSeqlensSize, maxSeqLen, stream);
}
when image.height == resizedHeight && image.width == resizedWidth, it still enter
into resizeImage function, https://github.com/NVIDIA/TensorRT-Edge-LLM/blob/release/0.8.0/cpp/runtime/imageUtils.cpp#L104-L127
if (mode == InterpolationMode::kBICUBIC)
{
stbir_resize(image.data(), image.width, image.height, kINPUT_STRIDE_BYTES, resizedImage.data(), newWidth,
newHeight, kOUTPUT_STRIDE_BYTES, STBIR_RGB, STBIR_TYPE_UINT8, STBIR_EDGE_CLAMP, STBIR_FILTER_CATMULLROM);
}
it will generate wrong/unstable output even if each input is fixed, the STBIR_FILTER_CATMULLROM mode can cause data fluctuations.
So how to fix ?
ref follow :
if (doResize)
{
auto [resizedHeight, resizedWidth] = getResizedImageSize(image.height, image.width);
if (resizedHeight != image.height || resizedWidth != image.width) {
rt::imageUtils::resizeImage(
image, mResizedImageHost, resizedWidth, resizedHeight, rt::imageUtils::InterpolationMode::kBICUBIC);
formatPatch(mResizedImageHost, imageGridTHWs, imageTokenLengths, cuSeqlensData, cuSeqlensSize,
maxSeqLen, stream);
}
else {
formatPatch(image, imageGridTHWs, imageTokenLengths, cuSeqlensData, cuSeqlensSize, maxSeqLen, stream);
}
}
else
{
formatPatch(image, imageGridTHWs, imageTokenLengths, cuSeqlensData, cuSeqlensSize, maxSeqLen, stream);
}
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start in cpp/multimodal/qwenViTRunner.cpp around the resizeImage call, then inspect cpp/runtime/imageUtils.cpp and its STBIR path. Ensure equal input and target dimensions do not go through the bicubic resize, while other cases retain the existing flow; done means the original image is formatted when dimensions match and resized output is used otherwise.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- computer-vision
- Issue type
- Bug
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Activity status
- Quiet
- Clarity
- Clearly specified
- Newbie friendliness
- 68/100