Skip to content

Evaluate perception(Camera)#

The performance of Autoware's recognition function (perception) is evaluated by calculating mAP (mean Average Precision) and other indices from the recognition results.

Run the perception module and pass the output perception topic to the evaluation library for evaluation.

Preparation#

In perception evaluation, machine-learning pre-trained models are used. If the model is not prepared in advance, Autoware will not output recognition results. If no evaluation results are produced, check to see if this has been done correctly.

Downloading Model Files#

Models are downloaded during Autoware setup. The method of downloading models depends on the version of Autoware you are using, so check which method is used. The following patterns exist.

Download with ansible#

When you run the ansible setup script, you will see Download artifacts? [y/N], type y and press enter (Autoware foundation's main branch use this method) https://github.com/autowarefoundation/autoware/blob/main/ansible/roles/artifacts/tasks/main.yaml

Automatically downloaded when the package is built#

If you are using a slightly older Autoware.universe, this is the one to use, until the commit hash of 13b96ad3c636389b32fea3a47dfb7cfb7813cadc. tensorrt_yolox/CMakeList.txt

Conversion of model files#

The downloaded onnx file is not to be used as-is, but to be converted to a TensorRT engine file for use. A conversion command is available, so source the autoware workspace and execute the command.

Let's assume that autoware is installed in $HOME/autoware.

source $HOME/autoware/install/setup.bash
ros2 launch tensorrt_yolox yolox.launch.xml use_decompress:=false build_only:=true

When the conversion command finishes, the engine file is output. The output destination changes according to the model download method, so check that the output is in the appropriate directory.

Download with ansible#

The following file is output.

$HOME/autoware_data/tensorrt_yolox/yolox-sPlus-T4-960x960-pseudo-finetune.EntropyV2-int8-batch1.engine

Automatic download at package build time#

The following file is output.

$HOME/autoware/install/tensorrt_yolox/share/tensorrt_yolox/data/yolox-sPlus-T4-960x960-pseudo-finetune.EntropyV2-int8-batch1.engine

(For evaluation on a single PC) modify the launch file#

To evaluate on a single PC, it is necessary to modify launch to output the recognition results of the camera. Change launch as follows.

❯ vcs diff src/
.................................
diff --git a/launch/tier4_perception_launch/launch/object_recognition/detection/camera_lidar_fusion_based_detection.launch.xml b/launch/tier4_perception_launch/launch/object_recognition/detection/camera_lidar_fusion_based_detection.launch.xml
index 9ca8ea3df..a35e8d00f 100644
--- a/launch/tier4_perception_launch/launch/object_recognition/detection/camera_lidar_fusion_based_detection.launch.xml
+++ b/launch/tier4_perception_launch/launch/object_recognition/detection/camera_lidar_fusion_based_detection.launch.xml
@@ -30,6 +30,14 @@
   <arg name="remove_unknown" default="true"/>
   <arg name="trust_distance" default="30.0"/>

+  <group>
+    <include file="$(find-pkg-share tensorrt_yolox)/launch/yolox.launch.xml" />
+  </group>
+
+  <group>
+    <include file="$(find-pkg-share bytetrack)/launch/bytetrack.launch.xml" />
+  </group>
+
   <!-- Jetson AGX -->
   <!-- <include file="$(find-pkg-share tensorrt_yolo)/launch/yolo.launch.xml">
     <arg name="image_raw0" value="$(var image_raw0)"/>
diff --git a/launch/tier4_perception_launch/launch/perception.launch.xml b/launch/tier4_perception_launch/launch/perception.launch.xml
index 0a2ef57f6..9a9b06379 100644
--- a/launch/tier4_perception_launch/launch/perception.launch.xml
+++ b/launch/tier4_perception_launch/launch/perception.launch.xml
@@ -33,7 +33,7 @@
   <arg name="camera_info6" default="/sensing/camera/camera6/camera_info"/>
   <arg name="image_raw7" default="/sensing/camera/camera7/image_rect_color"/>
   <arg name="camera_info7" default="/sensing/camera/camera7/camera_info"/>
-  <arg name="image_number" default="6" description="choose image raw number(0-7)"/>
+  <arg name="image_number" default="1" description="choose image raw number(0-7)"/>
   <arg name="use_vector_map" default="true" description="use vector map in prediction"/>
   <arg name="use_pointcloud_map" default="true" description="use pointcloud map in detection"/>
   <arg name="use_object_filter" default="true" description="use object filter"/>
diff --git a/perception/tensorrt_yolox/launch/yolox.launch.xml b/perception/tensorrt_yolox/launch/yolox.launch.xml
index b697b1f50..b9cb53102 100644
--- a/perception/tensorrt_yolox/launch/yolox.launch.xml
+++ b/perception/tensorrt_yolox/launch/yolox.launch.xml
@@ -1,7 +1,7 @@
 <?xml version="1.0"?>
 <launch>
   <arg name="input/image" default="/sensing/camera/camera0/image_rect_color"/>
-  <arg name="output/objects" default="/perception/object_recognition/detection/rois0"/>
+  <arg name="output/objects_yolox" default="/perception/object_recognition/detection/rois0"/>
   <arg name="model_name" default="yolox-tiny"/>
   <arg name="model_path" default="$(find-pkg-share tensorrt_yolox)/data"/>
   <arg name="score_threshold" default="0.35"/>
@@ -16,7 +16,7 @@

   <node pkg="tensorrt_yolox" exec="tensorrt_yolox_node_exe" name="tensorrt_yolox" output="screen">
     <remap from="~/in/image" to="$(var input/image)"/>
-    <remap from="~/out/objects" to="$(var output/objects)"/>
+    <remap from="~/out/objects" to="$(var output/objects_yolox)"/>
     <param name="score_threshold" value="$(var score_threshold)"/>
     <param name="nms_threshold" value="$(var nms_threshold)"/>
     <param name="model_path" value="$(var model_path)/$(var model_name).onnx"/>

Evaluation method#

The perception_2d evaluation is executed by launching the perception_2d.launch.py file. Launching the file executes the following steps:

  1. Execute launch of evaluation node (perception_2d_evaluator_node), logging_simulator.launch file and ros2 bag play command
  2. Autoware receives sensor data output from input rosbag and outputs camera, and the perception module performs recognition.
  3. The evaluation node subscribes to /perception/object_recognition/detection{/detected}/rois{camera_no} and evaluates data. The result is dumped into a file.
  4. When the playback of the rosbag is finished, Autoware's launch is automatically terminated, and the evaluation is completed.

Evaluation results#

The results are calculated for each subscription. The format and available states are described below.

Perception Normal#

When the following conditions are satisfied by executing the evaluation function of perception_eval

  1. frame_result.pass_fail_result contains at least one object (tp_object_results ! = [] and fp_object_results ! = [] and fn_objects ! = [])
  2. no object fail (frame_result.pass_fail_result.get_fail_object_num() == 0)

Perception Error#

The perception evaluation output is marked as Error when condition for Normal is not met.

Topic name and data type used by evaluation node#

Subscribed topics:

Topic name Data type
/perception/object_recognition/detection/rois{camera_no} tier4_perception_msgs/msg/DetectedObjectsWithFeature
/perception/object_recognition/detection/tracked/rois{camera_no} tier4_perception_msgs/msg/DetectedObjectsWithFeature

Published topics:

Topic name Data type
- -

Arguments passed to logging_simulator.launch#

To make Autoware processing less resource-consuming, modules that are not relevant to evaluation are disabled by passing the false parameter as a launch argument. The following parameters are set to false:

  • localization: false
  • planning: false
  • control: false
  • sensing: false / true (default value is false. Specify by LaunchSensing key for each t4_dataset in the scenario)
  • perception_mode: camera_lidar_fusion

NOTE: The tf in the bag is used to align the localization during annotation and simulation. Therefore, localization is invalid.

Dependent libraries#

The perception evaluation step bases on the perception_eval library.

Division of roles of driving_log_replayer with dependent libraries#

driving_log_replayer package is in charge of the connection with ROS. The actual perception evaluation is conducted in perception_eval library. The perception_eval is a ROS-independent library, it cannot receive ROS objects. Also, ROS timestamps use nanoseconds while the t4_dataset format is based on milliseconds (because it uses nuScenes), so the values must be properly converted before using the library's functions.

driving_log_replayer subscribes the topic output from the perception module of Autoware, converts it to the data format defined in perception_eval, and passes it on. It is also responsible for publishing and visualizing the evaluation results from perception_eval on proper ROS topic.

perception_eval is in charge of the part that compares the detection results passed from driving_log_replayer with ground truth data, calculates the index, and outputs the results.

About simulation#

State the information required to run the simulation.

Topic to be included in the input rosbag#

Must contain the required topics in t4_dataset format.

The vehicle's ECU CAN and sensors data topics are required for the evaluation to be run correctly. The following example shows the topic list available in evaluation input rosbag when multiple LiDARs and Cameras are used in a real-world vehicle configuration.

/sensing/lidar/concatenated/pointcloud is used if the scenario LaunchSensing: false.

If there is more than one CAMERA, include all on-board camera_info and image_rect_color_compressed.

Topic name Data type
/pacmod/from_can_bus can_msgs/msg/Frame
/sensing/camera/camera*/camera_info sensor_msgs/msg/CameraInfo
/sensing/camera/camera*/image_rect_color/compressed sensor_msgs/msg/CompressedImage
/sensing/gnss/ublox/fix_velocity geometry_msgs/msg/TwistWithCovarianceStamped
/sensing/gnss/ublox/nav_sat_fix sensor_msgs/msg/NavSatFix
/sensing/gnss/ublox/navpvt ublox_msgs/msg/NavPVT
/sensing/imu/tamagawa/imu_raw sensor_msgs/msg/Imu
/sensing/lidar/concatenated/pointcloud sensor_msgs/msg/PointCloud2
/sensing/lidar/*/velodyne_packets velodyne_msgs/VelodyneScan
/tf tf2_msgs/msg/TFMessage

The vehicle topics can be included instead of CAN.

Topic name Data type
/sensing/camera/camera*/camera_info sensor_msgs/msg/CameraInfo
/sensing/camera/camera*/image_rect_color/compressed sensor_msgs/msg/CompressedImage
/sensing/gnss/ublox/fix_velocity geometry_msgs/msg/TwistWithCovarianceStamped
/sensing/gnss/ublox/nav_sat_fix sensor_msgs/msg/NavSatFix
/sensing/gnss/ublox/navpvt ublox_msgs/msg/NavPVT
/sensing/imu/tamagawa/imu_raw sensor_msgs/msg/Imu
/sensing/lidar/concatenated/pointcloud sensor_msgs/msg/PointCloud2
/sensing/lidar/*/velodyne_packets velodyne_msgs/VelodyneScan
/tf tf2_msgs/msg/TFMessage
/vehicle/status/control_mode autoware_vehicle_msgs/msg/ControlModeReport
/vehicle/status/gear_status autoware_vehicle_msgs/msg/GearReport
/vehicle/status/steering_status autoware_vehicle_msgs/SteeringReport
/vehicle/status/turn_indicators_status autoware_vehicle_msgs/msg/TurnIndicatorsReport
/vehicle/status/velocity_status autoware_vehicle_msgs/msg/VelocityReport

Topics that must not be included in the input rosbag#

Topic name Data type
/clock rosgraph_msgs/msg/Clock

The clock is output by the --clock option of ros2 bag play, so if it is recorded in the bag itself, it is output twice, so it is not included in the bag.

About Evaluation#

State the information necessary for the evaluation.

Scenario Format#

There are two types of evaluation: use case evaluation and database evaluation. Use case evaluation is performed on a single dataset, while database evaluation uses multiple datasets and takes the average of the results for each dataset.

In the database evaluation, the vehicle_id should be able to be set for each data set, since the calibration values may change. Also, it is necessary to set whether or not to activate the sensing module.

See sample.

Evaluation Result Format#

See sample.

The evaluation results by perception_eval under the conditions specified in the scenario are output for each frame. Only the final line has a different format from the other lines since the final metrics are calculated after all data has been flushed.

The format of each frame and the metrics format are shown below. NOTE: common part of the result file format, which has already been explained, is omitted.

Format of each frame:

{
  "Frame": {
    "CameraType": "Evaluated camera",
    "FrameName": "Frame number of t4_dataset used for evaluation",
    "FrameSkip": "Number of times that an object was requested to be evaluated but the evaluation was skipped because there was no ground truth in the dataset within 75msec",
    "PassFail": {
      "Result": { "Total": "Success or Fail", "Frame": "Success or Fail" },
      "Info": {
        "TP": "Number of TPs",
        "FP": "Number of FPs",
        "FN": "Number of FNs"
      }
    }
  }
}

Metrics Data Format:

{
  "Frame": {
    "FinalScore": {
      "Score": {
        "TP": {
          "ALL": "TP rate for all labels",
          "label0": "TP rate of label0",
          "label1": "TP rate of label1"
        },
        "FP": {
          "ALL": "FP rate for all labels",
          "label0": "FP rate of label0",
          "label1": "FP rate of label1"
        },
        "FN": {
          "ALL": "FN rate for all labels",
          "label0": "FN rate of label0",
          "label1": "FN rate of label1"
        },
        "TN": {
          "ALL": "TN rate for all labels",
          "label0": "TN rate of label0",
          "label1": "TN rate of label1"
        },
        "AP(Center Distance)": {
          "ALL": "AP(Center Distance) rate for all labels",
          "label0": "AP(Center Distance) rate of label0",
          "label1": "AP(Center Distance) rate of label1"
        },
        "APH(Center Distance)": {
          "ALL": "APH(Center Distance) rate for all labels",
          "label0": "APH(Center Distance) rate of label0",
          "label1": "APH(Center Distance) rate of label1"
        },
        "AP(IoU 2D)": {
          "ALL": "AP(IoU 2D) rate for all labels",
          "label0": "AP(IoU 2D) rate of label0",
          "label1": "AP(IoU 2D) rate of label1"
        },
        "APH(IoU 2D)": {
          "ALL": "APH(IoU 2D) rate for all labels",
          "label0": "APH(IoU 2D) rate of label0",
          "label1": "APH(IoU 2D) rate of label1"
        }
      },
      "ConfusionMatrix": {
        "label0(GroundTruth)": {
          "label0(Prediction)": "value",
          "label1(Prediction)": "value"
        },
        "label1(GroundTruth)": {
          "label0(Prediction)": "value",
          "label1(Prediction)": "value"
        }
      }
    }
  }
}

pickle file#

In database evaluation, it is necessary to replay multiple rosbags, but due to the ROS specification, it is impossible to use multiple bags in a single launch. Since one rosbag, i.e., one t4_dataset, requires one launch, it is necessary to execute as many launches as the number of datasets contained in the database evaluation.

Since database evaluation cannot be done in a single launch, perception outputs a file scene_result.pkl in addition to result.jsonl file. A pickle file is a python object saved as a file, PerceptionEvaluationManager.frame_results of perception_eval. The dataset evaluation can be performed by reading all the objects recorded in the pickle file and outputting the index of the dataset's average.

Result file of database evaluation#

In the case of a database evaluation with multiple datasets in the scenario, a file named database_result.json is output to the results directory.

The format is the same as the Metrics Data Format.