Games202作业5（完结）

单帧降噪

也就是针对图像空间进行降噪

也就是我们需要在像素（i,j）的四周进行采样，然后将采样的权重加到一起，然后所有的权重和像素的乘积也加到一起，然后相除，得到最终滤波后的(i,j)像素期望的结果。

Buffer2D<Float3> Denoiser::Filter(const FrameInfo &frameInfo) {int height = frameInfo.m_beauty.m_height;int width = frameInfo.m_beauty.m_width;Buffer2D<Float3> filteredImage = CreateBuffer2D<Float3>(width, height);int kernelRadius = 16;#pragma omp parallel forfor (int y = 0; y < height; y++) {for (int x = 0; x < width; x++) {float sum_of_weight = 0.0;Float3 sum_of_weight_values(0.0);Float3 position_i = frameInfo.m_position(x, y);Float3 color_i = frameInfo.m_beauty(x, y);Float3 normal_i = frameInfo.m_normal(x, y);int boundary_x_max = std::min(kernelRadius + x, width);int boundary_x_min = std::max(0, x - kernelRadius);int boundary_y_max = std::min(kernelRadius + y, height);int boundary_y_min = std::max(y - kernelRadius, 0);for(int j_y = boundary_y_min; j_y < boundary_y_max; j_y++){for(int j_x = boundary_x_min; j_x < boundary_x_max; j_x++){Float3 position_j = frameInfo.m_position(j_x, j_y);Float3 color_j = frameInfo.m_beauty(j_x, j_y);Float3 normal_j = frameInfo.m_normal(j_x, j_y);//positionfloat p_w = - Dot(position_i - position_j, position_i - position_j) / (2.0 * m_sigmaCoord * m_sigmaCoord);//colorfloat c_w = - Dot(color_i - color_j, color_i - color_j) / (2.0 * m_sigmaColor * m_sigmaColor);//normalfloat n_w = - SafeAcos(Dot(normal_i, normal_j)) * SafeAcos(Dot(normal_i, normal_j)) / (2.0 * m_sigmaNormal * m_sigmaNormal);//planefloat pl_w = - (Dot(normal_i,(position_j - position_i) / std::max(SafeSqrt(Dot(position_j - position_i, position_j - position_i)), 0.0001f)) * Dot(normal_i,(position_j - position_i) / std::max(SafeSqrt(Dot(position_j - position_i, position_j - position_i)), 0.0001f))) / (2.0 * m_sigmaPlane * m_sigmaPlane);float weight = std::exp(p_w + c_w + n_w + pl_w);sum_of_weight += weight;sum_of_weight_values += (frameInfo.m_beauty(j_x, j_y) * weight);}}// TODO: Joint bilateral filterfilteredImage(x, y) = sum_of_weight_values / std::max(0.0001f,sum_of_weight);}}return filteredImage;
}

结果

滤波前：

非常大的噪声

滤波后：

非常的夸张了

投影到上一帧

首先我们需要求出motion vector来判断上一帧的像素位置来补充当前帧的对应像素的信息

根据当前帧的当前像素的物体的编号求当前帧的像素对应物体的model矩阵
求该物体上一帧的model矩阵
求上一帧的screen坐标
判断上一帧的screen坐标最终渲染出来的物体是不是该物体，不是就continue
将上一帧的结果保存在m_accColor中

void Denoiser::Reprojection(const FrameInfo &frameInfo) {int height = m_accColor.m_height;int width = m_accColor.m_width;Matrix4x4 preWorldToScreen =m_preFrameInfo.m_matrix[m_preFrameInfo.m_matrix.size() - 1];//P矩阵Matrix4x4 preWorldToCamera =m_preFrameInfo.m_matrix[m_preFrameInfo.m_matrix.size() - 2];//V矩阵#pragma omp parallel forfor (int y = 0; y < height; y++) {for (int x = 0; x < width; x++) {m_valid(x, y) = false;m_misc(x, y) = Float3(0.f);auto current_position = frameInfo.m_position(x, y);int current_id = frameInfo.m_id(x, y);if(current_id == -1){continue;}Matrix4x4 currentModel = frameInfo.m_matrix[current_id];Matrix4x4 lastModel = m_preFrameInfo.m_matrix[current_id];Matrix4x4 currentModelInv = Inverse(currentModel);Float3 lastScreen = preWorldToScreen(preWorldToCamera(currentModelInv(current_position, Float3::Point), Float3::Point), Float3::Point);if(lastScreen.x <0 || lastScreen.x >= width || lastScreen.y < 0 || lastScreen.y >= height){continue;}int actualId = m_preFrameInfo.m_id(lastScreen.x, lastScreen.y);if(actualId == current_id){m_valid(x, y) = true;m_misc(x, y) = m_accColor(lastScreen.x, lastScreen.y);}// TODO: Reproject}}std::swap(m_misc, m_accColor);
}

累积多帧

首先我们需要计算均值和方差

Float3 mean(0.0);
for(int i = x_min; i < x_max; i++){for(int j = y_min; j < y_max; j++){mean += curFilteredColor(i, j);}
}
mean /= std::max(static_cast<float>((x_max - x_min) * (y_max - y_min)), 0.0001f);
Float3 variance(0.0);
for(int i = x_min; i < x_max; i++){for(int j = y_min; j < y_max; j++){variance += ((curFilteredColor(i, j) - mean) * (curFilteredColor(i, j) - mean));}
}
variance /= std::max(static_cast<float>((x_max - x_min) * (y_max - y_min)), 0.0001f);
Float3 sigma = SafeSqrt(variance);

然后混合计算：

void Denoiser::TemporalAccumulation(const Buffer2D<Float3> &curFilteredColor) {int height = m_accColor.m_height;int width = m_accColor.m_width;int kernelRadius = 3;
#pragma omp parallel forfor (int y = 0; y < height; y++) {for (int x = 0; x < width; x++) {// TODO: Temporal clampint x_min = std::max(0, x - kernelRadius);int x_max = std::min(width, x + kernelRadius);int y_min = std::max(0, y - kernelRadius);int y_max = std::min(height, y + kernelRadius);Float3 mean(0.0);for(int i = x_min; i < x_max; i++){for(int j = y_min; j < y_max; j++){mean += curFilteredColor(i, j);}}mean /= std::max(static_cast<float>((x_max - x_min) * (y_max - y_min)), 0.0001f);Float3 variance(0.0);for(int i = x_min; i < x_max; i++){for(int j = y_min; j < y_max; j++){variance += ((curFilteredColor(i, j) - mean) * (curFilteredColor(i, j) - mean));}}variance /= std::max(static_cast<float>((x_max - x_min) * (y_max - y_min)), 0.0001f);Float3 sigma = SafeSqrt(variance);Float3 color = m_accColor(x, y);Float3 color_last = Clamp(color, mean - sigma * m_colorBoxK, mean + sigma * m_colorBoxK);float alpha = 1.0f;if(m_valid(x, y)){alpha = m_alpha;}// TODO: Exponential moving averagem_misc(x, y) = Lerp(color_last, curFilteredColor(x, y), alpha);}}std::swap(m_misc, m_accColor);
}

用A-Trous Wavelet加速单帧降噪

其实就是膨胀卷积原理，间隔采样

Buffer2D<Float3> Denoiser::Valelt_Filter(const FrameInfo &frameInfo) {int height = frameInfo.m_beauty.m_height;int width = frameInfo.m_beauty.m_width;Buffer2D<Float3> filteredImage = CreateBuffer2D<Float3>(width, height);int kernelRadius = 16;
#pragma omp parallel forfor (int y = 0; y < height; y++) {for (int x = 0; x < width; x++) {float sum_of_weight = 0.0;Float3 sum_of_weight_values(0.0);Float3 position_i = frameInfo.m_position(x, y);Float3 color_i = frameInfo.m_beauty(x, y);Float3 normal_i = frameInfo.m_normal(x, y);int boundary_x_max = std::min(kernelRadius + x, width);int boundary_x_min = std::max(0, x - kernelRadius);int boundary_y_max = std::min(kernelRadius + y, height);int boundary_y_min = std::max(y - kernelRadius, 0);for(int pass = 0; pass < 5; pass++){for(int i_left = -2; i_left <=2 ;i_left++){for(int j_left = -2; j_left <= 2; j_left ++){int j_x = x + i_left * std::pow(pass,2);int j_y = y + j_left * std::pow(pass,2);Float3 position_j = frameInfo.m_position(j_x, j_y);Float3 color_j = frameInfo.m_beauty(j_x, j_y);Float3 normal_j = frameInfo.m_normal(j_x, j_y);//positionfloat p_w = - Dot(position_i - position_j, position_i - position_j) / (2.0 * m_sigmaCoord * m_sigmaCoord);//colorfloat c_w = - Dot(color_i - color_j, color_i - color_j) / (2.0 * m_sigmaColor * m_sigmaColor);//normalfloat n_w = - SafeAcos(Dot(normal_i, normal_j)) * SafeAcos(Dot(normal_i, normal_j)) / (2.0 * m_sigmaNormal * m_sigmaNormal);//planefloat pl_w = - (Dot(normal_i,(position_j - position_i) / std::max(SafeSqrt(Dot(position_j - position_i, position_j - position_i)), 0.0001f)) * Dot(normal_i,(position_j - position_i) / std::max(SafeSqrt(Dot(position_j - position_i, position_j - position_i)), 0.0001f))) / (2.0 * m_sigmaPlane * m_sigmaPlane);float weight = std::exp(p_w + c_w + n_w + pl_w);sum_of_weight += weight;sum_of_weight_values += (frameInfo.m_beauty(j_x, j_y) * weight);}}}// TODO: Joint bilateral filterfilteredImage(x, y) = sum_of_weight_values / std::max(sum_of_weight, 0.0001f);}}return filteredImage;
}

使用A-Trous Wavelet之后：