revert: "feat(lidar_transfusion): update TransFusion-L model" (#1478)

Revert "feat(lidar_transfusion): update TransFusion-L model"

Author: shmpwk

perception/lidar_transfusion/CMakeLists.txt

@@ -137,58 +137,6 @@ if(TRT_AVAIL AND CUDA_AVAIL AND CUDNN_AVAIL)
   if(BUILD_TESTING)
     find_package(ament_lint_auto REQUIRED)
     ament_lint_auto_find_test_dependencies()
-
-    find_package(ament_cmake_gtest REQUIRED)
-    ament_auto_add_gtest(test_detection_class_remapper
-      test/test_detection_class_remapper.cpp
-    )
-    ament_auto_add_gtest(test_ros_utils
-      test/test_ros_utils.cpp
-    )
-    ament_auto_add_gtest(test_nms
-      test/test_nms.cpp
-    )
-
-    # Temporary disabled, tracked by:
-    # https://github.com/autowarefoundation/autoware.universe/issues/7724
-    # ament_auto_add_gtest(test_voxel_generator
-    #   test/test_voxel_generator.cpp
-    # )
-
-    # # preprocess kernel test
-    # add_executable(test_preprocess_kernel
-    #   test/test_preprocess_kernel.cpp
-    # )
-    # target_include_directories(test_preprocess_kernel PUBLIC
-    #   ${test_preprocess_kernel_SOURCE_DIR}
-    # )
-    # target_link_libraries(test_preprocess_kernel
-    #   transfusion_cuda_lib
-    #   gtest
-    #   gtest_main
-    # )
-    # ament_add_test(test_preprocess_kernel
-    #   GENERATE_RESULT_FOR_RETURN_CODE_ZERO
-    #   COMMAND "$<TARGET_FILE:test_preprocess_kernel>"
-    # )
-
-    # # postprocess kernel test
-    # add_executable(test_postprocess_kernel
-    #   test/test_postprocess_kernel.cpp
-    # )
-    # target_include_directories(test_postprocess_kernel PUBLIC
-    #   ${test_postprocess_kernel_SOURCE_DIR}
-    # )
-    # target_link_libraries(test_postprocess_kernel
-    #   transfusion_cuda_lib
-    #   gtest
-    #   gtest_main
-    # )
-    # ament_add_test(test_postprocess_kernel
-    #   GENERATE_RESULT_FOR_RETURN_CODE_ZERO
-    #   COMMAND "$<TARGET_FILE:test_postprocess_kernel>"
-    # )
-
   endif()
 
   ament_auto_package(

perception/lidar_transfusion/README.md

@@ -59,31 +59,13 @@ ros2 launch lidar_transfusion lidar_transfusion.launch.xml log_level:=debug
 
 ## Assumptions / Known limits
 
-This library operates on raw cloud data (bytes). It is assumed that the input pointcloud message has following format:
-
-```python
-[
-  sensor_msgs.msg.PointField(name='x', offset=0, datatype=7, count=1),
-  sensor_msgs.msg.PointField(name='y', offset=4, datatype=7, count=1),
-  sensor_msgs.msg.PointField(name='z', offset=8, datatype=7, count=1),
-  sensor_msgs.msg.PointField(name='intensity', offset=12, datatype=2, count=1)
-]
-```
-
-This input may consist of other fields as well - shown format is required minimum.
-For debug purposes, you can validate your pointcloud topic using simple command:
-
-```bash
-ros2 topic echo <input_topic> --field fields
-```
-
 ## Trained Models
 
 You can download the onnx format of trained models by clicking on the links below.
 
-- TransFusion: [transfusion.onnx](https://awf.ml.dev.web.auto/perception/models/transfusion/t4xx1_90m/v2/transfusion.onnx)
+- TransFusion: [transfusion.onnx](https://awf.ml.dev.web.auto/perception/models/transfusion/v1/transfusion.onnx)
 
-The model was trained in TIER IV's internal database (~11k lidar frames) for 50 epochs.
+The model was trained in TIER IV's internal database (~11k lidar frames) for 20 epochs.
 
 ### Changelog
 

perception/lidar_transfusion/config/transfusion.param.yaml

@@ -4,9 +4,8 @@
     class_names: ["CAR", "TRUCK", "BUS", "BICYCLE", "PEDESTRIAN"]
     trt_precision: fp16
     voxels_num: [5000, 30000, 60000]  # [min, opt, max]
-    point_cloud_range: [-92.16, -92.16, -3.0, 92.16, 92.16, 7.0]  # [x_min, y_min, z_min, x_max, y_max, z_max]
-    voxel_size: [0.24, 0.24, 10.0]  # [x, y, z]
-    num_proposals: 500
+    point_cloud_range: [-76.8, -76.8, -3.0, 76.8, 76.8, 5.0]  # [x_min, y_min, z_min, x_max, y_max, z_max]
+    voxel_size: [0.3, 0.3, 8.0]  # [x, y, z]
     onnx_path: "$(var model_path)/transfusion.onnx"
     engine_path: "$(var model_path)/transfusion.engine"
     # pre-process params
@@ -18,4 +17,4 @@
     iou_nms_search_distance_2d: 10.0
     iou_nms_threshold: 0.1
     yaw_norm_thresholds: [0.3, 0.3, 0.3, 0.3, 0.0]  # refers to the class_names
-    score_threshold: 0.1
+    score_threshold: 0.2

perception/lidar_transfusion/include/lidar_transfusion/preprocess/pointcloud_densification.hpp

@@ -51,7 +51,7 @@ class DensificationParam
 
 struct PointCloudWithTransform
 {
-  cuda::unique_ptr<uint8_t[]> data_d{nullptr};
+  cuda::unique_ptr<float[]> points_d{nullptr};
   std_msgs::msg::Header header;
   size_t num_points{0};
   Eigen::Affine3f affine_past2world;

perception/lidar_transfusion/include/lidar_transfusion/preprocess/preprocess_kernel.hpp

@@ -56,8 +56,12 @@ class PreprocessCuda
     unsigned int * mask, float * voxels, unsigned int * pillar_num, float * voxel_features,
     unsigned int * voxel_num, unsigned int * voxel_idxs);
 
+  // cudaError_t generateVoxelsInput_launch(
+  //   uint8_t * cloud_data, CloudInfo & cloud_info, unsigned int points_agg, unsigned int
+  //   points_size, float time_lag, float * affine_transform, float * points);
+
   cudaError_t generateSweepPoints_launch(
-    const uint8_t * input_data, size_t points_size, int input_point_step, float time_lag,
+    const float * input_points, size_t points_size, int input_point_step, float time_lag,
     const float * transform, float * output_points);
 
 private:

perception/lidar_transfusion/include/lidar_transfusion/preprocess/voxel_generator.hpp

@@ -18,17 +18,15 @@
 #include "lidar_transfusion/cuda_utils.hpp"
 #include "lidar_transfusion/preprocess/pointcloud_densification.hpp"
 #include "lidar_transfusion/preprocess/preprocess_kernel.hpp"
-#include "lidar_transfusion/ros_utils.hpp"
 #include "lidar_transfusion/transfusion_config.hpp"
+#include "lidar_transfusion/utils.hpp"
 
 #ifdef ROS_DISTRO_GALACTIC
 #include <tf2_eigen/tf2_eigen.h>
 #else
 #include <tf2_eigen/tf2_eigen.hpp>
 #endif
 
-#include <autoware_point_types/types.hpp>
-
 #include <sensor_msgs/msg/point_cloud2.hpp>
 
 #include <memory>
@@ -38,8 +36,8 @@
 
 namespace lidar_transfusion
 {
-constexpr size_t AFF_MAT_SIZE = 16;  // 4x4 matrix
-constexpr size_t MAX_CLOUD_STEP_SIZE = sizeof(autoware_point_types::PointXYZIRCAEDT);
+constexpr size_t AFF_MAT_SIZE = 16;         // 4x4 matrix
+constexpr size_t MAX_CLOUD_STEP_SIZE = 35;  // PointXYZIRCADT
 
 class VoxelGenerator
 {

perception/lidar_transfusion/include/lidar_transfusion/ros_utils.hpp

@@ -17,84 +17,16 @@
 
 #include "lidar_transfusion/utils.hpp"
 
-#include <autoware_point_types/types.hpp>
-
 #include <autoware_perception_msgs/msg/detected_object_kinematics.hpp>
 #include <autoware_perception_msgs/msg/detected_objects.hpp>
 #include <autoware_perception_msgs/msg/object_classification.hpp>
 #include <autoware_perception_msgs/msg/shape.hpp>
-#include <sensor_msgs/msg/point_field.hpp>
 
-#include <cstdint>
 #include <string>
 #include <vector>
 
-#define CHECK_OFFSET(offset, structure, field) \
-  static_assert(                               \
-    offsetof(structure, field) == offset,      \
-    "Offset of " #field " in " #structure " does not match expected offset.")
-#define CHECK_TYPE(type, structure, field)                 \
-  static_assert(                                           \
-    std::is_same<decltype(structure::field), type>::value, \
-    "Type of " #field " in " #structure " does not match expected type.")
-#define CHECK_FIELD(offset, type, structure, field) \
-  CHECK_OFFSET(offset, structure, field);           \
-  CHECK_TYPE(type, structure, field)
-
 namespace lidar_transfusion
 {
-using sensor_msgs::msg::PointField;
-
-CHECK_FIELD(0, float, autoware_point_types::PointXYZIRCAEDT, x);
-CHECK_FIELD(4, float, autoware_point_types::PointXYZIRCAEDT, y);
-CHECK_FIELD(8, float, autoware_point_types::PointXYZIRCAEDT, z);
-CHECK_FIELD(12, uint8_t, autoware_point_types::PointXYZIRCAEDT, intensity);
-
-struct CloudInfo
-{
-  uint32_t x_offset{0};
-  uint32_t y_offset{sizeof(float)};
-  uint32_t z_offset{sizeof(float) * 2};
-  uint32_t intensity_offset{sizeof(float) * 3};
-  uint8_t x_datatype{PointField::FLOAT32};
-  uint8_t y_datatype{PointField::FLOAT32};
-  uint8_t z_datatype{PointField::FLOAT32};
-  uint8_t intensity_datatype{PointField::UINT8};
-  uint8_t x_count{1};
-  uint8_t y_count{1};
-  uint8_t z_count{1};
-  uint8_t intensity_count{1};
-  uint32_t point_step{sizeof(autoware_point_types::PointXYZIRCAEDT)};
-  bool is_bigendian{false};
-
-  bool operator!=(const CloudInfo & rhs) const
-  {
-    return x_offset != rhs.x_offset || y_offset != rhs.y_offset || z_offset != rhs.z_offset ||
-           intensity_offset != rhs.intensity_offset || x_datatype != rhs.x_datatype ||
-           y_datatype != rhs.y_datatype || z_datatype != rhs.z_datatype ||
-           intensity_datatype != rhs.intensity_datatype || x_count != rhs.x_count ||
-           y_count != rhs.y_count || z_count != rhs.z_count ||
-           intensity_count != rhs.intensity_count || is_bigendian != rhs.is_bigendian;
-  }
-
-  friend std::ostream & operator<<(std::ostream & os, const CloudInfo & info)
-  {
-    os << "x_offset: " << static_cast<int>(info.x_offset) << std::endl;
-    os << "y_offset: " << static_cast<int>(info.y_offset) << std::endl;
-    os << "z_offset: " << static_cast<int>(info.z_offset) << std::endl;
-    os << "intensity_offset: " << static_cast<int>(info.intensity_offset) << std::endl;
-    os << "x_datatype: " << static_cast<int>(info.x_datatype) << std::endl;
-    os << "y_datatype: " << static_cast<int>(info.y_datatype) << std::endl;
-    os << "z_datatype: " << static_cast<int>(info.z_datatype) << std::endl;
-    os << "intensity_datatype: " << static_cast<int>(info.intensity_datatype) << std::endl;
-    os << "x_count: " << static_cast<int>(info.x_count) << std::endl;
-    os << "y_count: " << static_cast<int>(info.y_count) << std::endl;
-    os << "z_count: " << static_cast<int>(info.z_count) << std::endl;
-    os << "intensity_count: " << static_cast<int>(info.intensity_count) << std::endl;
-    os << "is_bigendian: " << static_cast<int>(info.is_bigendian) << std::endl;
-    return os;
-  }
-};
 
 void box3DToDetectedObject(
   const Box3D & box3d, const std::vector<std::string> & class_names,

perception/lidar_transfusion/include/lidar_transfusion/transfusion_config.hpp

@@ -26,9 +26,8 @@ class TransfusionConfig
 public:
   TransfusionConfig(
     const std::vector<int64_t> & voxels_num, const std::vector<double> & point_cloud_range,
-    const std::vector<double> & voxel_size, const int num_proposals,
-    const float circle_nms_dist_threshold, const std::vector<double> & yaw_norm_thresholds,
-    const float score_threshold)
+    const std::vector<double> & voxel_size, const float circle_nms_dist_threshold,
+    const std::vector<double> & yaw_norm_thresholds, const float score_threshold)
   {
     if (voxels_num.size() == 3) {
       max_voxels_ = voxels_num[2];
@@ -62,9 +61,6 @@ class TransfusionConfig
       voxel_y_size_ = static_cast<float>(voxel_size[1]);
       voxel_z_size_ = static_cast<float>(voxel_size[2]);
     }
-    if (num_proposals > 0) {
-      num_proposals_ = num_proposals;
-    }
     if (score_threshold > 0.0) {
       score_threshold_ = score_threshold;
     }
@@ -80,9 +76,6 @@ class TransfusionConfig
     grid_x_size_ = static_cast<std::size_t>((max_x_range_ - min_x_range_) / voxel_x_size_);
     grid_y_size_ = static_cast<std::size_t>((max_y_range_ - min_y_range_) / voxel_y_size_);
     grid_z_size_ = static_cast<std::size_t>((max_z_range_ - min_z_range_) / voxel_z_size_);
-
-    feature_x_size_ = grid_x_size_ / out_size_factor_;
-    feature_y_size_ = grid_y_size_ / out_size_factor_;
   }
 
   ///// INPUT PARAMETERS /////
@@ -114,7 +107,7 @@ class TransfusionConfig
   const std::size_t out_size_factor_{4};
   const std::size_t max_num_points_per_pillar_{points_per_voxel_};
   const std::size_t num_point_values_{4};
-  std::size_t num_proposals_{200};
+  const std::size_t num_proposals_{200};
   // the number of feature maps for pillar scatter
   const std::size_t num_feature_scatter_{pillar_feature_size_};
   // the score threshold for classification

perception/lidar_transfusion/include/lidar_transfusion/utils.hpp

@@ -36,6 +36,52 @@ struct Box3D
   float yaw;
 };
 
+struct CloudInfo
+{
+  uint32_t x_offset;
+  uint32_t y_offset;
+  uint32_t z_offset;
+  uint32_t intensity_offset;
+  uint8_t x_datatype;
+  uint8_t y_datatype;
+  uint8_t z_datatype;
+  uint8_t intensity_datatype;
+  uint8_t x_count;
+  uint8_t y_count;
+  uint8_t z_count;
+  uint8_t intensity_count;
+  uint32_t point_step;
+  bool is_bigendian;
+
+  CloudInfo()
+  : x_offset(0),
+    y_offset(4),
+    z_offset(8),
+    intensity_offset(12),
+    x_datatype(7),
+    y_datatype(7),
+    z_datatype(7),
+    intensity_datatype(7),
+    x_count(1),
+    y_count(1),
+    z_count(1),
+    intensity_count(1),
+    point_step(16),
+    is_bigendian(false)
+  {
+  }
+
+  bool operator!=(const CloudInfo & rhs) const
+  {
+    return x_offset != rhs.x_offset || y_offset != rhs.y_offset || z_offset != rhs.z_offset ||
+           intensity_offset != rhs.intensity_offset || x_datatype != rhs.x_datatype ||
+           y_datatype != rhs.y_datatype || z_datatype != rhs.z_datatype ||
+           intensity_datatype != rhs.intensity_datatype || x_count != rhs.x_count ||
+           y_count != rhs.y_count || z_count != rhs.z_count ||
+           intensity_count != rhs.intensity_count || is_bigendian != rhs.is_bigendian;
+  }
+};
+
 enum NetworkIO { voxels = 0, num_points, coors, cls_score, dir_pred, bbox_pred, ENUM_SIZE };
 
 // cspell: ignore divup

perception/lidar_transfusion/lib/preprocess/pointcloud_densification.cpp

@@ -93,15 +93,16 @@ void PointCloudDensification::enqueue(
   affine_world2current_ = affine_world2current;
   current_timestamp_ = rclcpp::Time(msg.header.stamp).seconds();
 
-  auto data_d = cuda::make_unique<uint8_t[]>(
-    sizeof(uint8_t) * msg.width * msg.height * msg.point_step / sizeof(uint8_t));
+  assert(sizeof(uint8_t) * msg.width * msg.height * msg.point_step % sizeof(float) == 1);
+  auto points_d = cuda::make_unique<float[]>(
+    sizeof(uint8_t) * msg.width * msg.height * msg.point_step / sizeof(float));
 
   CHECK_CUDA_ERROR(cudaMemcpyAsync(
-    data_d.get(), msg.data.data(), sizeof(uint8_t) * msg.width * msg.height * msg.point_step,
+    points_d.get(), msg.data.data(), sizeof(uint8_t) * msg.width * msg.height * msg.point_step,
     cudaMemcpyHostToDevice, stream_));
 
   PointCloudWithTransform pointcloud = {
-    std::move(data_d), msg.header, msg.width * msg.height, affine_world2current.inverse()};
+    std::move(points_d), msg.header, msg.width * msg.height, affine_world2current.inverse()};
 
   pointcloud_cache_.push_front(std::move(pointcloud));
 }