Obstacel Detector: Fix inference input dimension

modules/perception/inference/onnx/libtorch_obstacle_detector.cc

@@ -32,33 +32,35 @@ LibtorchObstacleDetection::LibtorchObstacleDetection() : device_(torch::kCPU) {
 }
 
 bool LibtorchObstacleDetection::Evaluate(
-  const std::vector<std::vector<double>>& imageFrame) {
+  const std::vector<std::vector<std::vector<double>>>& imageFrame) {
   // Sanity checks.
   omp_set_num_threads(1);
   if (imageFrame.size() == 0) {
-    AINFO << "Got no channel in image frame!";
+    AERROR << "Got no channel in image frame!";
     return false;
   }
   if (imageFrame[0].size() == 0) {
-    AINFO << "Got no image frame in channel 0!";
+    AERROR << "Got no image frame in channel 0!";
     return false;
   }
-  if (imageFrame[0].size() != 72000) {
-    AINFO << "imageFrame[0].size() = " << imageFrame[0].size() << ", skiping!";
+  if (imageFrame[0].size() != 369664) {
+    AERROR << "imageFrame[0].size() = " << imageFrame[0].size() << ", skiping!";
     return false;
   }
-  // image imput size is 1920 * 1080 = 2073600
-  torch::Tensor image_tensor = torch::empty(32 * 3 * 3 * 3);
+  // image imput size is 608 * 608 = 369664
+  torch::Tensor image_tensor = torch::empty(1 * 3 * 608 * 608);
   float* data = image_tensor.data_ptr<float>();
 
   for (const auto& channel : imageFrame) {
     for (const auto& i : channel) {
-      *data++ = static_cast<float>(i) / 32767.0;
+      for (const auto& j : i) {
+        *data++ = static_cast<float>(j) / 255.0;
+      }
     }
   }
 
   torch::Tensor torch_input = torch::from_blob(image_tensor.data_ptr<float>(),
-                                               {32, 3, 3, 3, 3});
+                                               {1, 3, 608, 608});
   std::vector<torch::jit::IValue> torch_inputs;
   torch_inputs.push_back(torch_input.to(device_));
 

modules/perception/inference/onnx/libtorch_obstacle_detector.h

@@ -20,6 +20,10 @@
 #include "torch/script.h"
 #include "torch/torch.h"
 
+#include "opencv2/core/core.hpp"
+#include "opencv2/highgui/highgui.hpp"
+#include "opencv2/imgproc/imgproc.hpp"
+
 namespace apollo {
 namespace perception {
 namespace inference {
@@ -30,7 +34,7 @@ class LibtorchObstacleDetection {
 
   ~LibtorchObstacleDetection() = default;
 
-  bool Evaluate(const std::vector<std::vector<double>>& imageFrame);
+  bool Evaluate(const std::vector<std::vector<std::vector<double>>>& imageFrame);
 
  private:
   void LoadModel();

modules/perception/inference/onnx/libtorch_obstacle_detector_test.cc

@@ -32,8 +32,7 @@ class LibtorchObstacleDetectionTest : public ::testing::Test {
 };
 
 TEST_F(LibtorchObstacleDetectionTest, is_) {
-  std::vector<std::vector<double>> imageFrame(4,
-    (std::vector<double> (2073600, 0.01)));
+  std::vector<std::vector<std::vector<double>>>imageFrame(3, std::vector<std::vector<double>>(608, std::vector<double>(608, 0.01)));
   bool result = obstacle_detection_.Evaluate(imageFrame);
   EXPECT_EQ(result, false);
 }