fsoco

Formula Student Objects in Context Dataset for the Formula Student Driverless competitions

FSOCO

Formula Student Objects in Context

Deprecation Notice 06.08.’20

This project will not be actively maintained anymore.

The second iteration of the FSOCO dataset is already in beta testing.
This second iteration does not contain the data from the first one from the start, but migration of your datasets is possible, if you fulfill the new requirements - for most only small changes will be needed to achieve this. You can read into the reasoning behind this break between the first and second iteration.

If you’re interested in either joining the waiting list for the beta testing or being notified when the official start happens, drop us a line: fsoco.dataset@gmail.com

What?

Open-Source Dataset for Objects that need to be recognized during the dynamic disciplines of the Formula Student Driverless competitions.

Team Data Type Annotation Type # Data # Cones
municHMotorsport e.V. Color Images Darknet YOLO Format 3745 18697
Elbflorace e.V. Color Images Darknet YOLO Format 853 3791
SCUTRacing Color Images Darknet YOLO Format 792 5896
DHBW Engineering e.V. Color Images VOC 600 5794
StarkStrom Augsburg e.V Color Images Darknet YOLO Format 1120 7501
AMZ Racing Color Images Darknet YOLO Format 791 5685
ITU Racing Color Images Darknet YOLO Format 600 8241
Raceyard Color Images MM-Label Tool Format 600 9491
EUFS Color Images Darknet YOLO Format 1094 4594
Dimitris Martin Arampatzis Color Images Darknet YOLO Format 600 12333
Unicamp E-Racing Driverless Color Images Darknet YOLO Format 844 6699
KTHFS Color Images Darknet YOLO Format 630 3434
GETracing Dortmund Color Images Darknet YOLO Format 600 1454
Vermilion Racing Color Images Darknet YOLO Format 897 3422
DART Racing Color Images LabelMe Webtool 1658 22690
Fast Forest Color Images Darknet YOLO Format 603 9503
FS Team Tallinn Color Images Darknet YOLO Format 645 1650
BME Formula Racing Team Color Images Darknet YOLO Format 1763 27877
Driverless UPC Color Images VOC 776 3885
Global-Formula-Racing Color Images Coco 646 7146
High-Octane Motorsports e.V. Color Images Darknet YOLO Format 860 1895
GreenTeam Uni Stuttgart Color Images Darknet YOLO Format 827 4245
FS Team Weingarten Color Images Darknet YOLO Format 920 3220
Mit Dut Colab Color Images North American Standard 4452 55357
Wisconsin Autonomous Color Images Darknet YOLO Format 630 3338
e-gnition Hamburg Color Images Darknet YOLO Format 700 3685
eForce Driverless Color Images Darknet YOLO Format 600 4676
Arist.u.r.t.le. Color Images Darknet YOLO Format 3127 13234
Horsepower Hannover Color Images Darknet YOLO Format 4373 7941
Dynamis PRC Color Images Darknet YOLO Format 600 5894
FST Lisboa Color Images Darknet YOLO Format 655 4221
UPBracing Team e.V. Color Images Darknet YOLO Format 949 8570
Formula Electric Belgium Color Images Darknet YOLO Format 600 4874
STUBA Green Team Color Images Darknet YOLO Format 1996 3346
FaSTTUBe Color Images Darknet YOLO Format 605 2196
CURE Mannheim Color Images VOC 600 3052
E-Team Squadra Corse Color Images Darknet YOLO Format & COCO Format 600 782
Bauman Racing Team Color Images Darknet YOLO Format 600 7159
Lions Racing Braunschweig Color Images Darknet YOLO Format 644 9196
MMR Driverless Color Images Darknet YOLO Format 600 8498
UPM Racing Color Images Darknet YOLO Format 633 6014
Uvigo Motorsport Color Images Darknet YOLO Format 600 2352
TUfast Racing Team Color Images Darknet YOLO Format 783 13080
Scuderia Mensa Color Images Darknet YOLO Format 600 3423
FaSTDa Racing Color Images Darknet YOLO Format 600 5864

Annotation Types

Here you’ll find the definitions for all different annotation types appearing in the datasets. If you need the labels in another format, please look for the according script in the scripts folder or write one and share your solution - sharing is caring ;)

Darknet YOLO

Darknet uses normalized image dimensions for the labels and defines the regions-of-interest (ROI) by their class, mid-point, width and height

Darknet Bounding-Box

# darknet-label.txt

0 0.255078125 0.545833333333 0.02421875 0.0583333333333
0 0.41328125 0.613194444444 0.040625 0.081944444444
0 0.81015625 0.780555555556 0.0734375 0.15

[class index][mid_x][mid_y][width][height]

VOC

VOC is a xml based description format. A label will be similiar to:

<object>
    <name>yellow-cone</name>
    <pose>Unspecified</pose>
    <truncated>0</truncated>
    <difficult>0</difficult>
    <polygon>
        <x1>877</x1>
        <y1>571</y1>
        <x2>897</x2>
        <y2>528</y2>
        <x3>916</x3>
        <y3>576</y3>
    </polygon>
</object>

VOC can be converted to Darknet YOLO by using the script provided in this repo. DHBW Engineering used polygons for marking the cones. An example:

Example polygon image

Munich Labeling Tool (https://github.com/ddavid/MM-label-tool)

[# cones]

[minX][minY][maxX][maxY][labelname][dist_from_width][dist_from_height]

The position is given in absolute pixel values, the distance is calculated in metres.

There is a converter to Darknet YOLO in Scripts.

North American Standard

The North American Standard system makes use of single .csv files containing the labels. Each row, other than the first, represents a new image. Each column, other than the first two, represent a label for that particular image.

The label format adheres to the following convention: [X-top_left, Y-top_left, Image Height, Image Width]. Note that all dimensions are in absolute pixel values. The first two columns consist of the image name and optionally a link to a public server hosting the images for convenient downloading.

A parser for converting these .csv files to a python list can be found in the scripts folder. This particular dataset does not include class labels.

LabelMe Webtool

LabelMe Webtool is a online labeling tool developed from MIT CSAIL The following XML format is produced

<annotation>
    <filename>yellow-cone</filename>
    <folder>yellow-cone</folder>
    <source>
        <sourceImage>The MIT-CSAIL database of objects and scenes</sourceImage>
        <sourceAnnotation>LabelMe Webtool</sourceAnnotation>
    </source>
    <object>
        <name>orange_big</name>
        <deleted>0</deleted>
        <verified>0</verified>
        <occluded>no</occluded>
        <attributes/>
        <parts>
            <hasparts/>
            <ispartof/>
        </parts>
        <date>01-Jun-2018 17:53:36</date>
        <id>0</id>
        <type>bounding_box</type>
        <polygon>
            <username>anonymous</username>
            <pt><x>265</x><y>762</y></pt>
            <pt><x>339</x><y>762</y></pt>
            <pt><x>339</x><y>620</y></pt>
            <pt><x>265</x><y>620</y></pt>
        </polygon>
    </object>
    <object></object>
    ...
    <object></object>
    <imagesize>
        <nrows>1080</nrows>
        <ncols>1920</ncols>
    </imagesize>
</annotation>

More information can be found here

Who?

municHMotorsport e.V.

Elbflorace e.V.

SCUTRacing

DHBW Engineering e.V.

StarkStrom Augsburg e.V.

AMZ Racing

ITU Racing

Raceyard

EUFS

Unicamp E-racing

GETracing

Vermilion Racing

DART Racing

Fast Forest

FS Team Tallinn

BME Formula Racing Team

DriverlessUPC

GFR

HighOctaneMotorsports

Green Team

FS Team Weingarten

Formula Student Team Delft Driverless

Wisconsin Autonomous

e-gnition Hamburg

e-force Prague

Aristurtle

HorsePowerHannover

Dynamis

UPBracing

FEBelgium

STUBA Green Team

FaSTTUBe

CURE Mannheim

E-Team Squadra Corse

Bauman Racing Team

Lions Racing

mmr-driverless

UPM Racing

Uvigo Motorsport

TUfast

Scuderia Mensa

FaSTDa

*…

Why?

Open-Source Dataset to accelerate the development of (camera-based) solutions for Object Detection in the context of the Formula Student Driverless competitions. Collecting raw data and annotating it accordingly is important, but is not feasible to be done well enough by one team within the time constraints of the competition.