Merge pull request #41 from masadcv/euclidean-impl

Fix bug in local Euclidean distance calculation for fast implementation

FastGeodis/fastgeodis.h

@@ -46,7 +46,7 @@ torch::Tensor generalised_geodesic2d_cuda(
 torch::Tensor generalised_geodesic3d_cuda(
     const torch::Tensor &image,
     const torch::Tensor &mask,
-    const std::vector<float> &spacing,
+    std::vector<float> spacing,
     const float &v,
     const float &l_grad,
     const float &l_eucl,
@@ -64,7 +64,7 @@ torch::Tensor generalised_geodesic2d_cpu(
 torch::Tensor generalised_geodesic3d_cpu(
     const torch::Tensor &image,
     const torch::Tensor &mask,
-    const std::vector<float> &spacing,
+    std::vector<float> spacing,
     const float &v,
     const float &l_grad,
     const float &l_eucl,

FastGeodis/fastgeodis_cpu.cpp

@@ -226,7 +226,7 @@ void geodesic_frontback_pass_cpu(
             ld += float(std::abs(h_i - 1)) * spacing[1];
             ld += float(std::abs(w_i - 1)) * spacing[2];
 
-            local_dist[h_i * 3 + w_i] = ld;
+            local_dist[h_i * 3 + w_i] = sqrt(ld);
         }
     }
 
@@ -342,13 +342,16 @@ void geodesic_frontback_pass_cpu(
 torch::Tensor generalised_geodesic3d_cpu(
     const torch::Tensor &image, 
     const torch::Tensor &mask, 
-    const std::vector<float> &spacing, 
+    std::vector<float> spacing, 
     const float &v, 
     const float &l_grad, 
     const float &l_eucl, 
     const int &iterations
     )
 {
+    // square spacing with transform
+    std::transform(spacing.begin(), spacing.end(), spacing.begin(), spacing.begin(), std::multiplies<float>());
+
     torch::Tensor image_local = image.clone();
     torch::Tensor distance = v * mask.clone();
 

FastGeodis/fastgeodis_cuda.cu

@@ -517,7 +517,7 @@ void geodesic_frontback_pass_cuda(
             ld += float(std::abs(h_i - 1)) * spacing[1];
             ld += float(std::abs(w_i - 1)) * spacing[2];
 
-            local_dist[h_i * 3 + w_i] = ld;
+            local_dist[h_i * 3 + w_i] = sqrt(ld);
         }
     }
     // copy local distances to GPU __constant__ memory
@@ -582,13 +582,16 @@ void geodesic_frontback_pass_cuda(
 torch::Tensor generalised_geodesic3d_cuda(
     const torch::Tensor &image, 
     const torch::Tensor &mask, 
-    const std::vector<float> &spacing, 
+    std::vector<float> spacing, 
     const float &v, 
     const float &l_grad, 
     const float &l_eucl, 
     const int &iterations
     )
 {
+    // square spacing with transform
+    std::transform(spacing.begin(), spacing.end(), spacing.begin(), spacing.begin(), std::multiplies<float>());
+
     torch::Tensor image_local = image.clone();
     torch::Tensor distance = v * mask.clone();
 

setup.py

@@ -122,13 +122,13 @@ def get_extensions():
 
 setup(
     name="FastGeodis",
-    version="1.0.0",
+    version="1.0.1",
     description="Fast Implementation of Generalised Geodesic Distance Transform for CPU (OpenMP) and GPU (CUDA)",
     long_description=long_description,
     long_description_content_type="text/markdown",
     url="https://github.com/masadcv/FastGeodis",
     author="Muhammad Asad",
-    author_email="[email protected]",
+    author_email="[email protected]",
     license="BSD-3-Clause License",
     classifiers=[
         "License :: OSI Approved :: BSD License",