Cycles Denoising: Prevent overfitting when using a very low radius
For example, when using a radius of 1, only 9 pixels (due to weighting maybe even less) will be used, but the transform code may still decide to use a 5-dimensional (or even higher) fit. This causes severe overfitting and therefore weird pixel values. To avoid this, this commit limits the amount of dimensions to a third of the pixel number. For a radius of 3 or more, this doesn't change anything, but for 1 and 2 it can prevent fireflies and/or negative values being produced.
This commit is contained in:
parent
3dee1f079f
commit
6cd1d34dc1
Notes:
blender-bot
2023-02-14 06:56:53 +01:00
Referenced by issue #51583, Denoising produces a lot of artifacts when a low neighbor weighting value is used
|
@ -37,6 +37,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
max(rect.y, y - radius));
|
||||
int2 high = make_int2(min(rect.z, x + radius + 1),
|
||||
min(rect.w, y + radius + 1));
|
||||
int num_pixels = (high.y - low.y) * (high.x - low.x);
|
||||
|
||||
/* === Shift feature passes to have mean 0. === */
|
||||
float feature_means[DENOISE_FEATURES];
|
||||
|
@ -46,8 +47,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
math_vector_add(feature_means, features, DENOISE_FEATURES);
|
||||
} END_FOR_PIXEL_WINDOW
|
||||
|
||||
float pixel_scale = 1.0f / ((high.y - low.y) * (high.x - low.x));
|
||||
math_vector_scale(feature_means, pixel_scale, DENOISE_FEATURES);
|
||||
math_vector_scale(feature_means, 1.0f / num_pixels, DENOISE_FEATURES);
|
||||
|
||||
/* === Scale the shifted feature passes to a range of [-1; 1], will be baked into the transform later. === */
|
||||
float *feature_scale = tempvector;
|
||||
|
@ -73,6 +73,8 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
|
||||
math_matrix_jacobi_eigendecomposition(feature_matrix, transform, DENOISE_FEATURES, 1);
|
||||
*rank = 0;
|
||||
/* Prevent overfitting when a small window is used. */
|
||||
int max_rank = min(DENOISE_FEATURES, num_pixels/3);
|
||||
if(pca_threshold < 0.0f) {
|
||||
float threshold_energy = 0.0f;
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++) {
|
||||
|
@ -81,7 +83,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
threshold_energy *= 1.0f - (-pca_threshold);
|
||||
|
||||
float reduced_energy = 0.0f;
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
|
||||
for(int i = 0; i < max_rank; i++, (*rank)++) {
|
||||
if(i >= 2 && reduced_energy >= threshold_energy)
|
||||
break;
|
||||
float s = feature_matrix[i*DENOISE_FEATURES+i];
|
||||
|
@ -89,7 +91,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
}
|
||||
}
|
||||
else {
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
|
||||
for(int i = 0; i < max_rank; i++, (*rank)++) {
|
||||
float s = feature_matrix[i*DENOISE_FEATURES+i];
|
||||
if(i >= 2 && sqrtf(s) < pca_threshold)
|
||||
break;
|
||||
|
|
|
@ -38,6 +38,7 @@ ccl_device void kernel_filter_construct_transform(const ccl_global float *ccl_re
|
|||
max(rect.y, y - radius));
|
||||
int2 high = make_int2(min(rect.z, x + radius + 1),
|
||||
min(rect.w, y + radius + 1));
|
||||
int num_pixels = (high.y - low.y) * (high.x - low.x);
|
||||
const ccl_global float *ccl_restrict pixel_buffer;
|
||||
int2 pixel;
|
||||
|
||||
|
@ -52,8 +53,7 @@ ccl_device void kernel_filter_construct_transform(const ccl_global float *ccl_re
|
|||
math_vector_add(feature_means, features, DENOISE_FEATURES);
|
||||
} END_FOR_PIXEL_WINDOW
|
||||
|
||||
float pixel_scale = 1.0f / ((high.y - low.y) * (high.x - low.x));
|
||||
math_vector_scale(feature_means, pixel_scale, DENOISE_FEATURES);
|
||||
math_vector_scale(feature_means, 1.0f / num_pixels, DENOISE_FEATURES);
|
||||
|
||||
/* === Scale the shifted feature passes to a range of [-1; 1], will be baked into the transform later. === */
|
||||
float feature_scale[DENOISE_FEATURES];
|
||||
|
@ -81,6 +81,8 @@ ccl_device void kernel_filter_construct_transform(const ccl_global float *ccl_re
|
|||
|
||||
math_matrix_jacobi_eigendecomposition(feature_matrix, transform, DENOISE_FEATURES, transform_stride);
|
||||
*rank = 0;
|
||||
/* Prevent overfitting when a small window is used. */
|
||||
int max_rank = min(DENOISE_FEATURES, num_pixels/3);
|
||||
if(pca_threshold < 0.0f) {
|
||||
float threshold_energy = 0.0f;
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++) {
|
||||
|
@ -89,7 +91,7 @@ ccl_device void kernel_filter_construct_transform(const ccl_global float *ccl_re
|
|||
threshold_energy *= 1.0f - (-pca_threshold);
|
||||
|
||||
float reduced_energy = 0.0f;
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
|
||||
for(int i = 0; i < max_rank; i++, (*rank)++) {
|
||||
if(i >= 2 && reduced_energy >= threshold_energy)
|
||||
break;
|
||||
float s = feature_matrix[i*DENOISE_FEATURES+i];
|
||||
|
@ -97,7 +99,7 @@ ccl_device void kernel_filter_construct_transform(const ccl_global float *ccl_re
|
|||
}
|
||||
}
|
||||
else {
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
|
||||
for(int i = 0; i < max_rank; i++, (*rank)++) {
|
||||
float s = feature_matrix[i*DENOISE_FEATURES+i];
|
||||
if(i >= 2 && sqrtf(s) < pca_threshold)
|
||||
break;
|
||||
|
|
|
@ -32,6 +32,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
max(rect.y, y - radius));
|
||||
int2 high = make_int2(min(rect.z, x + radius + 1),
|
||||
min(rect.w, y + radius + 1));
|
||||
int num_pixels = (high.y - low.y) * (high.x - low.x);
|
||||
|
||||
__m128 feature_means[DENOISE_FEATURES];
|
||||
math_vector_zero_sse(feature_means, DENOISE_FEATURES);
|
||||
|
@ -40,7 +41,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
math_vector_add_sse(feature_means, DENOISE_FEATURES, features);
|
||||
} END_FOR_PIXEL_WINDOW_SSE
|
||||
|
||||
__m128 pixel_scale = _mm_set1_ps(1.0f / ((high.y - low.y) * (high.x - low.x)));
|
||||
__m128 pixel_scale = _mm_set1_ps(1.0f / num_pixels);
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++) {
|
||||
feature_means[i] = _mm_mul_ps(_mm_hsum_ps(feature_means[i]), pixel_scale);
|
||||
}
|
||||
|
@ -68,6 +69,8 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
math_matrix_jacobi_eigendecomposition(feature_matrix, transform, DENOISE_FEATURES, 1);
|
||||
|
||||
*rank = 0;
|
||||
/* Prevent overfitting when a small window is used. */
|
||||
int max_rank = min(DENOISE_FEATURES, num_pixels/3);
|
||||
if(pca_threshold < 0.0f) {
|
||||
float threshold_energy = 0.0f;
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++) {
|
||||
|
@ -76,7 +79,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
threshold_energy *= 1.0f - (-pca_threshold);
|
||||
|
||||
float reduced_energy = 0.0f;
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
|
||||
for(int i = 0; i < max_rank; i++, (*rank)++) {
|
||||
if(i >= 2 && reduced_energy >= threshold_energy)
|
||||
break;
|
||||
float s = feature_matrix[i*DENOISE_FEATURES+i];
|
||||
|
@ -84,7 +87,7 @@ ccl_device void kernel_filter_construct_transform(const float *ccl_restrict buff
|
|||
}
|
||||
}
|
||||
else {
|
||||
for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
|
||||
for(int i = 0; i < max_rank; i++, (*rank)++) {
|
||||
float s = feature_matrix[i*DENOISE_FEATURES+i];
|
||||
if(i >= 2 && sqrtf(s) < pca_threshold)
|
||||
break;
|
||||
|
|
Loading…
Reference in New Issue