Open Perception Lidar Model Training Service

Overview

Open Perception Lidar Model Training Service is a cloud-based service to train perception lidar model using pointpillars algorithm from your data, to better detect obstacles in your environment.

Prerequisites

• Apollo 6.0 or higher version.

• Baidu Cloud BOS service registered according to document

• Fuel service account on Apollo Dreamland

Main Steps

• Data collection

• Job submission

• Model training result

Data Collection

Data Recording

Collecting sensor data from lidar and cameras in different scenarios covering your autonomous driving environment as much as possible, please make sure the scenarios have different types of obstacles such as pedestrians and vehicles. Then labeling the sensor data using kitti data format.

Data format

• We use Kitti data format as training data format:

    INPUT_DATA_PATH:
        training:
            calib
            image_2
            label_2
            velodyne
        testing:
            calib
            image_2
            velodyne
        train.txt
	val.txt
        trainval.txt
        test.txt 

• Supported obstacle detection categories:

    bus, Car, construction_vehicle, Truck, barrier, Cyclist, motorcycle, Pedestrian, traffic_cone

When labeling your data, `type` must be one of the above categories (please note the uppercase).

Job Submission

Upload data to BOS

Requirements of the folder structure for job submission：

1. Input Data Path: upload your data to INPUT_DATA_PATH directory.

2. Output Data Path: if the model is trained successfully, an onnx file will be saved to the OUTPUT_DATA_PATH directory.

Submit job on Dreamland

Go to Apollo Dreamland, login with Baidu account, choose Apollo Fuel --> Jobs，New Job, Perception Lidar Model Training，and input the correct BOS path as in Upload data to BOS section.

Model Training Result

• Once a job is done, you should be expecting one email per job including the results and Model Path.