gaussian_kernel_2d()

const std::vector< float > & MayaFlux::Yantra::VisionGpuExecutor::gaussian_kernel_2d ( uint32_t  radius, float  sigma  )

private

2D Gaussian kernel for convolution, cached by (radius, sigma bit pattern).

Sigma is a tuning parameter that rarely changes frame to frame; recomputing exp() over (2*radius+1)^2 taps and reallocating the kernel every call is pure repeated work for an identical result.

Parameters

radius	Radius of the kernel in pixels. Kernel size is (2*radius + 1)^2.
sigma	Standard deviation of the Gaussian.

Returns

Normalized kernel weights as a flat vector in row-major order.

Definition at line 10 of file VisionImageDispatch.cpp.

{
    static std::unordered_map<uint64_t, std::vector<float>> cache;
    const uint64_t key = (static_cast<uint64_t>(std::bit_cast<uint32_t>(sigma)) << 32)
        | radius;
    auto it = cache.find(key);
    if (it != cache.end())
        return it->second;
 
    const uint32_t diam = 2 * radius + 1;
    std::vector<float> k(static_cast<size_t>(diam) * diam);
    float sum = 0.0F;
    for (uint32_t y = 0; y < diam; ++y) {
        for (uint32_t x = 0; x < diam; ++x) {
            const float fx = static_cast<float>(x) - static_cast<float>(radius);
            const float fy = static_cast<float>(y) - static_cast<float>(radius);
            const float v = std::exp(-(fx * fx + fy * fy) / (2.0F * sigma * sigma));
            k[y * diam + x] = v;
            sum += v;
        }
    }
    for (auto& v : k)
        v /= sum;
 
    return cache.emplace(key, std::move(k)).first->second;
}

References radius.

Referenced by op_canny(), and op_harris_response().

Here is the caller graph for this function: