Skip to content

Commit

Permalink
[hdr] sampling: minor code cleanup
Browse files Browse the repository at this point in the history
  • Loading branch information
@fabiencastan
fabiencastan committed Jul 21, 2020
1 parent 084d8f5 commit c469129
Show file tree
Hide file tree
Showing 2 changed files with 56 additions and 42 deletions.
95 changes: 55 additions & 40 deletions src/aliceVision/hdr/sampling.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -152,9 +152,10 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
const float area = float(diameter * diameter);

std::vector<std::pair<int, int>> vec_blocks;
for(int cy = 0; cy < imageHeight; cy += params.blockSize - diameter)
const auto step = params.blockSize - diameter;
for(int cy = 0; cy < imageHeight; cy += step)
{
for(int cx = 0; cx < imageWidth; cx += params.blockSize - diameter)
for(int cx = 0; cx < imageWidth; cx += step)
{
vec_blocks.push_back(std::make_pair(cx, cy));
}
Expand All @@ -171,8 +172,7 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
// Load image
readImage(imagePaths[idBracket], img, colorspace);

if(img.Width() != imageWidth ||
img.Height() != imageHeight)
if(img.Width() != imageWidth || img.Height() != imageHeight)
{
std::stringstream ss;
ss << "Failed to extract samples, the images with multi-bracketing do not have the same image resolution.\n"
Expand Down Expand Up @@ -202,26 +202,26 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
integral(imgIntegral, blockInput);
integral(imgIntegralSquare, imgSquare);

for(int i = radiusp1; i < imgIntegral.Height() - params.radius; ++i)
for(int y = radiusp1; y < imgIntegral.Height() - params.radius; ++y)
{
for(int j = radiusp1; j < imgIntegral.Width() - params.radius; ++j)
for(int x = radiusp1; x < imgIntegral.Width() - params.radius; ++x)
{
image::Rgb<double> S1 = imgIntegral(i + params.radius, j + params.radius) + imgIntegral(i - radiusp1, j - radiusp1) - imgIntegral(i + params.radius, j - radiusp1) - imgIntegral(i - radiusp1, j + params.radius);
image::Rgb<double> S2 = imgIntegralSquare(i + params.radius, j + params.radius) + imgIntegralSquare(i - radiusp1, j - radiusp1) - imgIntegralSquare(i + params.radius, j - radiusp1) - imgIntegralSquare(i - radiusp1, j + params.radius);
image::Rgb<double> S1 = imgIntegral(y + params.radius, x + params.radius) + imgIntegral(y - radiusp1, x - radiusp1) - imgIntegral(y + params.radius, x - radiusp1) - imgIntegral(y - radiusp1, x + params.radius);
image::Rgb<double> S2 = imgIntegralSquare(y + params.radius, x + params.radius) + imgIntegralSquare(y - radiusp1, x - radiusp1) - imgIntegralSquare(y + params.radius, x - radiusp1) - imgIntegralSquare(y - radiusp1, x + params.radius);

PixelDescription pd;

pd.exposure = exposure;
pd.mean.r() = blockInput(i,j).r();
pd.mean.g() = blockInput(i,j).g();
pd.mean.b() = blockInput(i,j).b();
pd.mean.r() = blockInput(y,x).r();
pd.mean.g() = blockInput(y,x).g();
pd.mean.b() = blockInput(y,x).b();
pd.variance.r() = (S2.r() - (S1.r()*S1.r()) / area) / area;
pd.variance.g() = (S2.g() - (S1.g()*S1.g()) / area) / area;
pd.variance.b() = (S2.b() - (S1.b()*S1.b()) / area) / area;

blockOutput(i, j).x = cx + j;
blockOutput(i, j).y = cy + i;
blockOutput(i, j).descriptions.push_back(pd);
blockOutput(y, x).x = cx + x;
blockOutput(y, x).y = cy + y;
blockOutput(y, x).descriptions.push_back(pd);
}
}
}
Expand All @@ -235,11 +235,11 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c

// Create samples image
#pragma omp parallel for
for (int i = params.radius; i < samples.Height() - params.radius; ++i)
for (int y = params.radius; y < samples.Height() - params.radius; ++y)
{
for (int j = params.radius; j < samples.Width() - params.radius; ++j)
for (int x = params.radius; x < samples.Width() - params.radius; ++x)
{
ImageSample & sample = samples(i, j);
ImageSample & sample = samples(y, x);
if (sample.descriptions.size() < 2)
{
continue;
Expand All @@ -250,11 +250,12 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
// Make sure we don't have a patch with high variance on any bracket.
// If the variance is too high somewhere, ignore the whole coordinate samples
bool valid = true;
const float maxVariance = 0.05f;
for (int k = 0; k < sample.descriptions.size(); ++k)
{
if (sample.descriptions[k].variance.r() > 0.05 ||
sample.descriptions[k].variance.g() > 0.05 ||
sample.descriptions[k].variance.b() > 0.05)
if (sample.descriptions[k].variance.r() > maxVariance ||
sample.descriptions[k].variance.g() > maxVariance ||
sample.descriptions[k].variance.b() > maxVariance)
{
valid = false;
break;
Expand All @@ -274,35 +275,46 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
{
bool valid = false;

if (sample.descriptions[k].mean.r() > 0.99f ||
sample.descriptions[k].mean.g() > 0.99f ||
sample.descriptions[k].mean.b() > 0.99f)
// Threshold on the max values, to avoid using fully saturated pixels
// TODO: on RAW images, values can be higher. May need to be computed dynamically?
const float maxValue = 0.99f;
if(sample.descriptions[k].mean.r() > maxValue ||
sample.descriptions[k].mean.g() > maxValue ||
sample.descriptions[k].mean.b() > maxValue)
{
continue;
}

if (sample.descriptions[k].mean.r() > 1.004f * sample.descriptions[k - 1].mean.r() ||
sample.descriptions[k].mean.g() > 1.004f * sample.descriptions[k - 1].mean.g() ||
sample.descriptions[k].mean.b() > 1.004f * sample.descriptions[k - 1].mean.b())
// Ensures that at least one channel is strictly increasing with increasing exposure
// TODO: check "exposure" params, we may have the same exposure multiple times
const float minIncreaseRatio = 1.004f;
if(sample.descriptions[k].mean.r() > minIncreaseRatio * sample.descriptions[k - 1].mean.r() ||
sample.descriptions[k].mean.g() > minIncreaseRatio * sample.descriptions[k - 1].mean.g() ||
sample.descriptions[k].mean.b() > minIncreaseRatio * sample.descriptions[k - 1].mean.b())
{
valid = true;
}

// Ensures that the values of each channel are increasing with increasing exposure
if (sample.descriptions[k].mean.r() < sample.descriptions[k - 1].mean.r() ||
sample.descriptions[k].mean.g() < sample.descriptions[k - 1].mean.g() ||
sample.descriptions[k].mean.b() < sample.descriptions[k - 1].mean.b())
{
valid = false;
}

if (sample.descriptions[k - 1].mean.norm() > 0.1f)
// If we have enough information to analyze the chrominance
const float minGlobalValue = 0.1f;
if(sample.descriptions[k - 1].mean.norm() > minGlobalValue)
{
// Check that both colors are similars
const float n1 = sample.descriptions[k - 1].mean.norm();
const float n2 = sample.descriptions[k].mean.norm();
const float dot = sample.descriptions[k - 1].mean.dot(sample.descriptions[k].mean);
const float cosa = dot / (n1*n2);
if (cosa < 0.95f)

const float maxCosa = 0.95f; // ~ 18deg
if(cosa < maxCosa)
{
valid = false;
}
Expand Down Expand Up @@ -353,13 +365,13 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
std::vector<Counters> counters_vec(omp_get_max_threads());

#pragma omp parallel for
for (int i = params.radius; i < samples.Height() - params.radius; ++i)
for (int y = params.radius; y < samples.Height() - params.radius; ++y)
{
Counters & counters_thread = counters_vec[omp_get_thread_num()];

for (int j = params.radius; j < samples.Width() - params.radius; ++j)
for (int x = params.radius; x < samples.Width() - params.radius; ++x)
{
ImageSample & sample = samples(i, j);
ImageSample & sample = samples(y, x);
UniqueDescriptor desc;

for (int k = 0; k < sample.descriptions.size(); ++k)
Expand Down Expand Up @@ -408,9 +420,9 @@ bool Sampling::extractSamplesFromImages(std::vector<ImageSample>& out_samples, c
item.second.resize(params.maxCountSample);
}

for (std::size_t l = 0; l < item.second.size(); ++l)
for (std::size_t i = 0; i < item.second.size(); ++i)
{
Coordinates coords = item.second[l];
Coordinates coords = item.second[i];

if (!samples(coords.second, coords.first).descriptions.empty())
{
Expand Down Expand Up @@ -454,7 +466,9 @@ void Sampling::analyzeSource(std::vector<ImageSample> & samples, int channelQuan

for (auto & item : _positions)
{
if (item.second.size() > 500)
// TODO: expose as parameters
const std::size_t maxSamples = 500;
if(item.second.size() > maxSamples)
{
// Shuffle and ignore the exceeding samples
std::random_shuffle(item.second.begin(), item.second.end());
Expand All @@ -465,6 +479,7 @@ void Sampling::analyzeSource(std::vector<ImageSample> & samples, int channelQuan

void Sampling::filter(size_t maxTotalPoints)
{
// TODO: avoid hardcoded value
size_t limitPerGroup = 510;
size_t total_points = maxTotalPoints + 1;

Expand All @@ -487,27 +502,27 @@ void Sampling::filter(size_t maxTotalPoints)
}
}

void Sampling::extractUsefulSamples(std::vector<ImageSample> & out_samples, std::vector<ImageSample> & samples, int imageIndex)
void Sampling::extractUsefulSamples(std::vector<ImageSample> & out_samples, const std::vector<ImageSample> & samples, int imageIndex) const
{
std::set<unsigned int> unique_indices;
std::set<unsigned int> uniqueIndices;

for (auto & item : _positions)
{
for (auto & pos : item.second)
{
if (pos.imageIndex == imageIndex)
{
unique_indices.insert(pos.sampleIndex);
uniqueIndices.insert(pos.sampleIndex);
}
}
}

for (auto & index : unique_indices)
// Export non-empty samples
for (auto & index : uniqueIndices)
{
if (samples[index].descriptions.size() > 0)
if (!samples[index].descriptions.empty())
{
out_samples.push_back(samples[index]);
samples[index].descriptions.clear();
}
}
}
Expand Down
3 changes: 1 addition & 2 deletions src/aliceVision/hdr/sampling.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,6 @@ std::ostream & operator<<(std::ostream& os, const PixelDescription & p);
std::istream & operator>>(std::istream& os, ImageSample & s);
std::istream & operator>>(std::istream& os, PixelDescription & p);

struct UniqueDescriptor;

class Sampling
{
Expand All @@ -65,7 +64,7 @@ class Sampling
public:
void analyzeSource(std::vector<ImageSample> & samples, int channelQuantization, int imageIndex);
void filter(size_t maxTotalPoints);
void extractUsefulSamples(std::vector<ImageSample> & out_samples, std::vector<ImageSample> & samples, int imageIndex);
void extractUsefulSamples(std::vector<ImageSample> & out_samples, const std::vector<ImageSample> & samples, int imageIndex) const;

static bool extractSamplesFromImages(std::vector<ImageSample>& out_samples, const std::vector<std::string> & imagePaths, const std::vector<float>& times, const size_t imageWidth, const size_t imageHeight, const size_t channelQuantization, const image::EImageColorSpace & colorspace, const Params params);

Expand Down

0 comments on commit c469129

Please sign in to comment.