GPU data augmentation
Decode one image, record five deterministic GPU augmentation views, and compose the source plus results through OA’s headless Vulkan renderer.
Source
| Surface | Repository path | SHA-256 |
|---|---|---|
| C++ | sdk/cpp/tutorials/vision/tuVisionDataAugmentation.cpp | 5da595a9424f3070 |
| Python | sdk/py/tutorials/vision/tuVisionDataAugmentation.py | 58f1a67ed1f69fc2 |
| Presentation | sdk/asset/docs/vision/oa-vision-data-augmentation.png | ea818dd5f205783c |
Example
Both programs apply the same flip, crop and resize, brightness and contrast, seeded Gaussian noise, clamp, and solarize operations. The semantic images remain GPU-resident until the composed figure reaches its terminal PNG output boundary.
1oa::Vector<oa::Image> views;2 views.reserve(6);3 views.pushBack(oa::move(decoded).getValue());4 const oa::Matrix& source = views.front().asMatrix();56 views.pushBack(rgbImage(oa::FnImage::flip(source, true, false)));7 auto crop = oa::FnImage::centerCrop(source, 280, 150);8 views.pushBack(rgbImage(oa::FnImage::resize(crop, 320, 180)));9 views.pushBack(rgbImage(10 oa::FnImage::brightnessContrast(source, 0.08F, 1.15F)));11 auto noisy = oa::FnImage::gaussianNoise(source, 0.0F, 0.035F, 2026U);12 views.pushBack(rgbImage(oa::FnImage::clamp(noisy, 0.0F, 1.0F)));13 views.pushBack(rgbImage(oa::FnImage::solarize(source, 0.55F, 1.0F)));14
This tutorial proves the transformation and residency contract. It does not claim a speedup over another image stack without matching decode policy, transforms, shapes, synchronization, and output validation.
Present
Six deterministic views. The source image and five GPU augmentation results are composed by the tutorial into this exact checked PNG.
present
Augmentation gallery
Source, horizontal flip, crop and resize, brightness and contrast, seeded Gaussian noise, and solarization.