bitfsd / FSACOCO

Open source cone detection dataset for FSAC

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

FSACOCO

Open-Source Dataset for Cones that need to be recognized during the dynamic disciplines of the Formula Student Autonomous China competition.

中文版文档见:README_cn.md

How to get the datasets

This dataset lives from your contribution. According to the current situation of FSD teams, we set up two methods to get the dataset.
1.Labeled Dataset: You first need to send your team's dataset to us. After we verify the validity of the dataset, we will send you all the current datasets we have. To promote the growth of the dataset, we have set a minimum contribution amount of 600 images.
2.Unlabeled Dataset: For some teams that have been established for a short time and do not meet the conditions of geting photo from real vehicle, they can send an e-mail application to obtain the unlabeled data provided by fsacoco for labeling, and then send back according to the acquisition method of labeled data. After verification, all data can be obtained.

At the same time, we also advocate that each FSD-team can extract your dataset(We provide a script called extract.py under scripts folder to extract photo from .bag file), and expand the unlabeled dataset in the format of one compressed package for every 800 images. The dataset should only include the photo in the running state (i.e. RES is go).

How to send your dataset

In order to solve the problem of transferring large files, we recommend https://airportal.cn (<2G) or https://pan.baidu.com/ (>2G) for file upload. Once the upload is completed, you need to manually send the download code to the e-mail address: bitfsd@163.com.

Annotation Types

MM-label-toos original labeled data and:

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-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

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

label tool

There is a labeling tool in tools. Small adaptations for labeling cones and additional functionality on BBox-Label-Tool.

Format:

[# 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.

Datasets Requirement

The label box must be closed to the cone as close as possible.
As shown in the figure below, the cones must meet the rules of FSAC:

Contributor

List of teams currently participating in dataset construction:

Beijing Institute of Technology Driverless Racing Team

FuZhou University Driverless Racing Team

Beihang University AERO Racing Team

Xihua University Driverless Racing Team

Hunan University Sliver Wing Team

Wuhan University of Science and Technology Chiji Racing Team

Hubei University of Automotive Technology Driverless Racing Team

Changchun University Driverless Racing Team

About

Open source cone detection dataset for FSAC


Languages

Language:Python 99.6%Language:Shell 0.4%