A web based labeling editor dedicated to the creation of training data for machine learning. The tool has been developed in the context of autonomous driving research. It supports images (.jpg and .png files) and point clouds (.pcd files). It is a Meteor app developed with React, Paper.js and three.js.
git clone https://github.com/Hitachi-Automotive-And-Industry-Lab/semantic-segmentation-editor
cd semantic-segmentation-editor
Change 'input-folder' and 'output-folder' according to your needs in settings.json
On Windows, use '/' separators, example "c:/Users/john/images"
meteor npm install
meteor --settings settings.json
Like any Meteor app, the editor will run by default on http://localhost:3000
(-p to change the port)
Check Meteor Environment Variables to configure your app
(MONGO_URL
, DISABLE_WEBSOCKETS
, etc...)
A Docker image is available at https://hub.docker.com/r/hitachiail/semantic-segmentation-editor/
To run it:
docker pull hitachiail/semantic-segmentation-editor
docker run -it -p PORT:3000 -v INPUT_FOLDER:/mnt/images -v OUTPUT_FOLDER:/mnt/pcd hitachiail/semantic-segmentation-editor:latest
Replace PORT
, INPUT_FOLDER
and OUTPUT_FOLDER
according to your needs.
Modifying this file let's you configure where are stored data of the app as well as the sets of classes available in the tool.
{
"configuration": {
"input-folder": "/mnt/images", // The root folder containing images and PCD files
"output-folder": "/mnt/pointcloud" // Segmentation data (only 3D) will be stored in this folder
},
// The different sets of classes available in the tool
// For object classes, only the 'label' field is mandatory
// The icon field can be set with an icon from the mdi-material-ui package
"sets-of-classes": [
{
"name": "Cityscapes", "objects": [
{"label": "VOID", "color": "#CFCFCF"},
{"label": "Road", "color": "#804080", "icon": "Road"},
{"label": "Sidewalk", "color": "#F423E8", "icon": "NaturePeople"},
{"label": "Parking", "color": "#FAAAA0", "icon": "Parking"},
{"label": "Rail Track", "color": "#E6968C", "icon": "Train"},
{"label": "Person", "color": "#DC143C", "icon": "Walk"},
{"label": "Rider", "color": "#FF0000", "icon": "Motorbike"},
{"label": "Car", "color": "#0000E8", "icon": "Car"}
},
{ ... }
]
}
The editor is built around 3 different screens:
The file navigator let's you browse available files to select a bitmap images or a point cloud for labeling
The bitmap image editor is dedicated to the labeling of jpg and png files by drawing polygons
The point cloud editor is dedicated to the labeling of point clouds by creating objects made of subsets of 3D points
There are several tools to create labeling polygons:
/api/listing
: List all annotated images/api/json/[PATH_TO_FILE]
: (2D only) Get the polygons and other data for that file/api/pcdtext/[PATH_TO_FILE]
: (3D only) Get the labeling of a pcd file using 2 addditional
columns: label
and object
/api/pcdfile/[PATH_TO_FILE]
: (3D only) The same but returned as "plain/text" attachment file download此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。