!6009 Add Affine transformation ut for Lite Dataset

Merge pull request !6009 from jiangzhiwen/lite/affine_ut

mindspore/ccsrc/minddata/dataset/kernels/image/lite_cv/image_process.cc

@@ -516,18 +516,10 @@ void ConvertBoxes(std::vector<std::vector<float>> &boxes, const std::vector<std:
 std::vector<int> ApplyNms(const std::vector<std::vector<float>> &all_boxes, std::vector<float> &all_scores, float thres,
                           int max_boxes) {
   int boxes_num = all_boxes.size();
-  std::vector<float> y1(boxes_num);
-  std::vector<float> x1(boxes_num);
-  std::vector<float> y2(boxes_num);
-  std::vector<float> x2(boxes_num);
   std::vector<float> areas(boxes_num);
   std::vector<int> order(boxes_num);
   for (int i = 0; i < boxes_num; i++) {
-    y1[i] = all_boxes[i][0];
-    x1[i] = all_boxes[i][1];
-    y2[i] = all_boxes[i][2];
-    x2[i] = all_boxes[i][3];
-    areas[i] = (x2[i] - x1[i] + 1) * (y2[i] - y1[i] + 1);
+    areas[i] = (all_boxes[i][3] - all_boxes[i][1] + 1) * (all_boxes[i][2] - all_boxes[i][0] + 1);
     order[i] = i;
   }
 
@@ -543,10 +535,10 @@ std::vector<int> ApplyNms(const std::vector<std::vector<float>> &all_boxes, std:
     int len = order.size() - 1;
     std::vector<float> ovr(len);
     for (int j = 0; j < len; j++) {
-      float xx1 = std::max(x1[i], x1[order[j + 1]]);
-      float yy1 = std::max(y1[i], y1[order[j + 1]]);
-      float xx2 = std::min(x2[i], x2[order[j + 1]]);
-      float yy2 = std::min(y2[i], y2[order[j + 1]]);
+      float xx1 = std::max(all_boxes[i][1], all_boxes[order[j + 1]][1]);
+      float yy1 = std::max(all_boxes[i][0], all_boxes[order[j + 1]][0]);
+      float xx2 = std::min(all_boxes[i][3], all_boxes[order[j + 1]][3]);
+      float yy2 = std::min(all_boxes[i][2], all_boxes[order[j + 1]][2]);
 
       float w = std::max(0.0f, xx2 - xx1 + 1);
       float h = std::max(0.0f, yy2 - yy1 + 1);
@@ -568,14 +560,15 @@ std::vector<int> ApplyNms(const std::vector<std::vector<float>> &all_boxes, std:
   return keep;
 }
 
-void WarpAffine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> dsize, UINT8_C3 borderValue) {
+template <typename Pixel_Type>
+void ImplementAffine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> &dsize, Pixel_Type borderValue) {
   double IM[6];
   for (int i = 0; i < 6; i++) {
     IM[i] = M[i];
   }
 
   double D = IM[0] * IM[4] - IM[1] * IM[3];
-  D = D != 0 ? 1. / D : 0;
+  D = D != 0 ? 1.0f / D : 0;
   double A11 = IM[4] * D, A22 = IM[0] * D;
   IM[0] = A11;
   IM[1] *= -D;
@@ -592,13 +585,22 @@ void WarpAffine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t>
       int src_x = IM[0] * x + IM[1] * y + IM[2];
       int src_y = IM[3] * x + IM[4] * y + IM[5];
       if (src_x >= 0 && src_y >= 0 && src_x < src.width_ && src_y < src.height_) {
-        UINT8_C3 src_pixel = static_cast<UINT8_C3 *>(src.data_ptr_)[src_y * src.width_ + src_x];
-        static_cast<UINT8_C3 *>(out_img.data_ptr_)[y * src.width_ + x] = src_pixel;
+        Pixel_Type src_pixel = static_cast<Pixel_Type *>(src.data_ptr_)[src_y * src.width_ + src_x];
+        static_cast<Pixel_Type *>(out_img.data_ptr_)[y * src.width_ + x] = src_pixel;
       } else {
-        static_cast<UINT8_C3 *>(out_img.data_ptr_)[y * src.width_ + x] = borderValue;
+        static_cast<Pixel_Type *>(out_img.data_ptr_)[y * src.width_ + x] = borderValue;
       }
     }
   }
 }
+
+void Affine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> dsize, UINT8_C1 borderValue) {
+  ImplementAffine(src, out_img, M, dsize, borderValue);
+}
+
+void Affine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> dsize, UINT8_C3 borderValue) {
+  ImplementAffine(src, out_img, M, dsize, borderValue);
+}
+
 }  // namespace dataset
 }  // namespace mindspore

mindspore/ccsrc/minddata/dataset/kernels/image/lite_cv/image_process.h

@@ -20,6 +20,7 @@
 #include <math.h>
 #include <vector>
 #include <algorithm>
+#include <iostream>
 
 #include "lite_cv/lite_mat.h"
 
@@ -68,13 +69,20 @@ bool SubStractMeanNormalize(const LiteMat &src, LiteMat &dst, float *mean, float
 bool Padd(const LiteMat &src, LiteMat &dst, int top, int bottom, int left, int right, PaddBorderType pad_type,
           uint8_t fill_b_or_gray, uint8_t fill_g, uint8_t fill_r);
 
-void WarpAffine(const LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> dsize, uint8_t borderValue[3]);
+/// \brief Apply affine transformation for 1 channel image
+void Affine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> dsize, UINT8_C1 borderValue);
 
+/// \brief Apply affine transformation for 3 channel image
+void Affine(LiteMat &src, LiteMat &out_img, double M[6], std::vector<size_t> dsize, UINT8_C3 borderValue);
+
+/// \brief Get default anchor boxes for Faster R-CNN, SSD, YOLO etc
 std::vector<std::vector<float>> GetDefaultBoxes(const BoxesConfig config);
 
+/// \brief Convert the prediction boxes to the actual boxes of (y, x, h, w)
 void ConvertBoxes(std::vector<std::vector<float>> &boxes, const std::vector<std::vector<float>> &default_boxes,
                   const BoxesConfig config);
 
+/// \brief Apply Non-Maximum Suppression
 std::vector<int> ApplyNms(const std::vector<std::vector<float>> &all_boxes, std::vector<float> &all_scores, float thres,
                           int max_boxes);
 

mindspore/ccsrc/minddata/dataset/kernels/image/lite_cv/lite_mat.h

@@ -27,17 +27,20 @@ namespace dataset {
 
 template <typename T>
 struct Chn1 {
+  Chn1(T c1) : c1(c1) {}
   T c1;
 };
 
 template <typename T>
 struct Chn2 {
+  Chn2(T c1, T c2) : c1(c1), c2(c2) {}
   T c1;
   T c2;
 };
 
 template <typename T>
 struct Chn3 {
+  Chn3(T c1, T c2, T c3) : c1(c1), c2(c2), c3(c3) {}
   T c1;
   T c2;
   T c3;
@@ -45,6 +48,7 @@ struct Chn3 {
 
 template <typename T>
 struct Chn4 {
+  Chn4(T c1, T c2, T c3, T c4) : c1(c1), c2(c2), c3(c3), c4(c4) {}
   T c1;
   T c2;
   T c3;

tests/ut/cpp/dataset/image_process_test.cc

@@ -21,6 +21,8 @@
 #include <opencv2/imgproc/types_c.h>
 #include "utils/log_adapter.h"
 
+#include <fstream>
+
 using namespace mindspore::dataset;
 class MindDataImageProcess : public UT::Common {
  public:
@@ -228,6 +230,7 @@ TEST_F(MindDataImageProcess, TestPadd) {
 }
 
 TEST_F(MindDataImageProcess, TestGetDefaultBoxes) {
+  std::string benchmark = "data/dataset/testLite/default_boxes.bin";
   BoxesConfig config;
   config.img_shape = {300, 300};
   config.num_default = {3, 6, 6, 6, 6, 6};
@@ -238,8 +241,25 @@ TEST_F(MindDataImageProcess, TestGetDefaultBoxes) {
   config.steps = {16, 32, 64, 100, 150, 300};
   config.prior_scaling = {0.1, 0.2};
 
+  int rows = 1917;
+  int cols = 4;
+  std::vector<double> benchmark_boxes(rows * cols);
+  std::ifstream in(benchmark, std::ios::in | std::ios::binary);
+  in.read(reinterpret_cast<char*>(benchmark_boxes.data()), benchmark_boxes.size() * sizeof(double));
+  in.close();
+
   std::vector<std::vector<float>> default_boxes = GetDefaultBoxes(config);
-  ASSERT_TRUE(default_boxes.size() == 1917);
+  EXPECT_EQ(default_boxes.size(), rows);
+  EXPECT_EQ(default_boxes[0].size(), cols);
+
+  double distance = 0.0f;
+  for (int i = 0; i < rows; i++) {
+    for (int j = 0; j < cols; j++) {
+      distance += pow(default_boxes[i][j] - benchmark_boxes[i * cols + j], 2);
+    }
+  }
+  distance = sqrt(distance);
+  EXPECT_LT(distance, 1e-5);
 }
 
 TEST_F(MindDataImageProcess, TestApplyNms) {
@@ -250,3 +270,67 @@ TEST_F(MindDataImageProcess, TestApplyNms) {
   ASSERT_TRUE(keep[1] == 0);
   ASSERT_TRUE(keep[2] == 1);
 }
+
+TEST_F(MindDataImageProcess, TestAffine) {
+  // The input matrix
+  // 0 0 1 0 0
+  // 0 0 1 0 0
+  // 2 2 3 2 2
+  // 0 0 1 0 0
+  // 0 0 1 0 0
+  size_t rows = 5;
+  size_t cols = 5;
+  LiteMat src(rows, cols);
+  for (size_t i = 0; i < rows; i++) {
+    for (size_t j = 0; j < cols; j++) {
+      if (i == 2 && j == 2) {
+        static_cast<UINT8_C1*>(src.data_ptr_)[i * cols + j] = 3;
+      } else if (i == 2) {
+        static_cast<UINT8_C1*>(src.data_ptr_)[i * cols + j] = 2;
+      } else if (j == 2) {
+        static_cast<UINT8_C1*>(src.data_ptr_)[i * cols + j] = 1;
+      } else {
+        static_cast<UINT8_C1*>(src.data_ptr_)[i * cols + j] = 0;
+      }
+    }
+  }
+
+  // Expect output matrix
+  // 0 0 2 0 0
+  // 0 0 2 0 0
+  // 1 1 3 1 1
+  // 0 0 2 0 0
+  // 0 0 2 0 0
+  LiteMat expect(rows, cols);
+  for (size_t i = 0; i < rows; i++) {
+    for (size_t j = 0; j < cols; j++) {
+      if (i == 2 && j == 2) {
+        static_cast<UINT8_C1*>(expect.data_ptr_)[i * cols + j] = 3;
+      } else if (i == 2) {
+        static_cast<UINT8_C1*>(expect.data_ptr_)[i * cols + j] = 1;
+      } else if (j == 2) {
+        static_cast<UINT8_C1*>(expect.data_ptr_)[i * cols + j] = 2;
+      } else {
+        static_cast<UINT8_C1*>(expect.data_ptr_)[i * cols + j] = 0;
+      }
+    }
+  }
+
+  double angle = 90.0f;
+  cv::Point2f center(rows / 2, cols / 2);
+  cv::Mat rotate_matrix = cv::getRotationMatrix2D(center, angle, 1.0);
+  double M[6];
+  for (size_t i = 0; i < 6; i++) {
+    M[i] = rotate_matrix.at<double>(i);
+  }
+  std::cout << std::endl;
+  LiteMat dst;
+  Affine(src, dst, M, {rows, cols}, UINT8_C1(0));
+
+  for (size_t i = 0; i < rows; i++) {
+    for (size_t j = 0; j < cols; j++) {
+      EXPECT_EQ(static_cast<UINT8_C1*>(expect.data_ptr_)[i * cols + j].c1,
+                static_cast<UINT8_C1*>(dst.data_ptr_)[i * cols + j].c1);
+    }
+  }
+}