Omniverse Replicator环境配置和使用说明

Omniverse Replicator使用说明

本教程将介绍Omniverse Replicator的环境配置和使用说明, 参加Sky Hackathon的同学可以参考本教程来合成训练数据集.

1. Omniverse环境配置

NVIDIA Omniverse 可以在任何搭载了 RTX 的设备上运行。为了获得理想性能，我们建议使用显存不低于 8GB 的 GeForce RTX 3070 或 NVIDIA RTX A4000 显卡。

组成部分	最低规格
支持的操作系统	Windows 11/ Windows 10（64 位 版本 1909 及更高版本）
CPU	Intel I7 /AMD Ryzen 2.5 GHz 或更高
CPU Core 核心数	4 个或更多
RAM	16Gb 或更多
存储	500Gb SSD 或更多
GPU	任何 RTX GPU
VRAM	6Gb 或更多
最低 视频驱动版本	单击 此处，查看新版驱动

Omniverse中的应用需要GPU和安装驱动, 如果您已安装则无需重复操作.

如果您未安装, 请参考:

• Windows: https://blog.csdn.net/kunhe0512/article/details/124331221
• Linux: https://blog.csdn.net/kunhe0512/article/details/125061911

您可能需要根据最新版本调整上面文章中的步骤

下面的步骤以Windows 10系统为例

1.a.安装Omniverse Launcher

1.a.1.在下面的地址下载Omniverse Launcher

https://www.nvidia.com/en-us/omniverse/download/

1.a.2.安装Omniverse Launcher

双击下载omniverse-launcher-win.exe, 按照要求安装.

1.a.3.登录

如果您没有账号, 请创建账号. 如果您已经有账号, 登录Omniverse Launcher

1.b 安装Omniverse CODE

在EXCHANGE中找到CODE, 点击并安装.

安装完之后, 在LIBRARY页面中选择CODE并加载

当出现以下页面, 意味着安装成功了

注意: 这里一定要等到RTX Loading字样消失在操作CODE, 否则可能会出现卡顿或者程序以外关闭

2. Omniverse Code介绍

Omniverse Code是我们生成图像数据集主要操作的工具, 我们使用的Replicator也是可以在里面操作的一个扩展库.

您可以在Omniverse Code里面输入准备好的代码, 然后生成您想创建的场景, 并通过Replicator合成训练数据集

下面是Omniverse Code的页面展示:

注意:在每次生成生成一个新的场景是, 需要点击左上角的File -> New -> Don’t Save. 连续更新代码, 然后点击Run(Ctrl + Enter)可能会出现程序意外关闭.

3.使用Omniverse Code/Replicator生成场景并合成图像数据集

1. 打开Omniverse Code
2. 在Script Editor里面输入以下代码:

# import replicator envirnoment 
import omni.replicator.core as rep

# setup random view range for camera: low point, high point
sequential_pos = [(-800, 220, -271),(800, 220,500)]

# position of look-at target
look_at_position = (-212, 78, 57)


# setup working layer 
with rep.new_layer():

# define 3d models: usd format file source link, class, initial position  
        WORKSHOP = 'http://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/ArchVis/Industrial/Buildings/Warehouse/Warehouse01.usd'
        CONVEYOR = 'http://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/DigitalTwin/Assets/Warehouse/Equipment/Conveyors/ConveyorBelt_A/ConveyorBelt_A23_PR_NVD_01.usd'
        BOX1     = 'http://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/ArchVis/Industrial/Containers/Cardboard/Cardbox_A3.usd'
        BOX2     = 'http://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/ArchVis/Industrial/Containers/Cardboard/Cardbox_B3.usd'
        BOX3     = 'http://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/ArchVis/Industrial/Containers/Cardboard/Cardbox_C3.usd'
        BOX4     = 'http://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/ArchVis/Industrial/Containers/Cardboard/Cardbox_D3.usd'
        workshop = rep.create.from_usd(WORKSHOP)
        conveyor1 = rep.create.from_usd(CONVEYOR)
        conveyor2 = rep.create.from_usd(CONVEYOR)
        box1 = rep.create.from_usd(BOX1,semantics=[('class', 'box')])
        box2 = rep.create.from_usd(BOX2,semantics=[('class', 'box')])
        box3 = rep.create.from_usd(BOX3,semantics=[('class', 'box')])
        box4 = rep.create.from_usd(BOX4,semantics=[('class', 'box')])
       
        with workshop:
            rep.modify.pose(
                position=(0,0,0),
                rotation=(0,-90,-90)
                )
        with conveyor1:
            rep.modify.pose(
                position=(-40,0,0),
                rotation=(0,-90,-90)
                )
        with conveyor2:
            rep.modify.pose(
                position=(-40,0,100),
                rotation=(-90,90,0)
                )        
                              
        with box1:
            rep.modify.pose(
                position=(-350,78,57),
                rotation=(0,-90,-90),
                scale=rep.distribution.uniform(1,1)
                )
        with box2:
            rep.modify.pose(
                position=(-100,78,57),
                rotation=(0,-90,-90),
                scale=rep.distribution.uniform(1,1)
                )   
        with box3:
            rep.modify.pose(
                position=(100,78,57),
                rotation=(0,-90,-90),
                scale=rep.distribution.uniform(1,1)
                )  
        with box4:
            rep.modify.pose(
                position=(200,78,57),
                rotation=(0,-90,-90),
                scale=rep.distribution.uniform(1,1)
                ) 

# define lighting function
        def sphere_lights(num):
               lights = rep.create.light(
                     light_type="Sphere",
                     temperature=rep.distribution.normal(3500, 500),
                     intensity=rep.distribution.normal(15000, 5000),
                     position=rep.distribution.uniform((-300, -300, -300), (300, 300, 300)),
                     scale=rep.distribution.uniform(50, 100),
                     count=num
               )
               return lights.node
        rep.randomizer.register(sphere_lights)

# define function to create random position range for target  
        def get_shapes():
            shapes = rep.get.prims(semantics=[('class', 'box')])
            with shapes:
                rep.modify.pose(
                    position=rep.distribution.uniform((0, -50, 0), (0, 50, 0)))
            return shapes.node
        rep.randomizer.register(get_shapes)

# Setup camera and attach it to render product
        camera = rep.create.camera(position=sequential_pos[0], look_at=look_at_position)
        render_product = rep.create.render_product(camera, resolution=(512, 512))

        with rep.trigger.on_frame(num_frames=100): #number of picture
               rep.randomizer.sphere_lights(4)    #number of lighting source 
               rep.randomizer.get_shapes()
               with camera:
                        rep.modify.pose(
                          position=rep.distribution.uniform(sequential_pos[0],sequential_pos[1]), look_at=look_at_position)

# Initialize and attach writer for Kitti format data 
        writer = rep.WriterRegistry.get("KittiWriter")
        writer.initialize(
                 output_dir="D:/sdg", 
                 bbox_height_threshold=25,
                 fully_visible_threshold=0.95,
                 omit_semantic_type=True
               )
        writer.attach([render_product])
        rep.orchestrator.preview()

3. 单击左下角的Run(Ctrl + Enter)按钮

4. 选择上面的’Replicator -> Start’

5. 在D:/sdg文件夹内查看生成的内容. 图片在Camera/rgb文件夹内, label把文件在Camera/object_detection文件夹内, 如下图所示
   

6. 完成上述步骤, 说明您已经可以成功使用Omniverse Replicator来合成数据