build details

Show: section status errors & todos local changes recent changes last change in-page changes feedback controls

Object Detection Dataset

Modified 2020-11-30 by Liam Paull

Download

Modified 2020-11-30 by Liam Paull

The dataset can be downloaded from here. We provide annotations and sample scripts for loading the annotations.

Overview

Modified 2021-10-31 by tanij

This dataset consists of 3 categories: traffic cones, duckies, and Duckiebots. All the dataset images were captured with Duckiebot cameras. We use a combination of images from the Duckietown logs database and our own captured logs. Images were captured in different lighting conditions, with different versions of Duckiebot models, and on different Duckietown maps. Below are some statistics and visualizations of our dataset:


Number of images  1956
Number of object categories  3
Number of objects annotated  5068


Category Details

Modified 2020-11-30 by Liam Paull

Traffic Cones

Modified 2020-11-30 by Liam Paull

Category name  cone
Number of instances  372
Category id  1

Duckies

Modified 2020-11-30 by Liam Paull

Category name  duckie
Number of instances  2570
Category id  2

Duckiebots

Modified 2021-10-31 by tanij

Category name  Duckiebot
Number of instances  2126
Category id  3
Number of old Duckiebot instances  1419
Number of new Duckiebot instances  707

Data Loading Scripts

Modified 2020-11-30 by Liam Paull

We provide some sample scripts for loading in the dataset here.

Data Collection Procedure

Modified 2021-10-31 by tanij

In this work, we first identify the most prominent objects that we see on the roads of Duckietown: duckies, Duckiebots and traffic cones. To begin our data collection procedure, we first identify all useful logs from the Duckietown logs database which contain the objects of interest. We then download and trim these logs so that the videos consist only of frames containing our objects of interest. Finally, we convert our videos to images (frames) while skipping some number of frames between each image to ensure that we get a diverse set of images.

In these logs, there are videos of older versions of Duckiebots with lots of wirings on them (DB17). However, new Duckiebots are much cleaner with only the battery visible. To ensure robust detections, we needed to capture this intra-class variation. Thus, we collected our own logs containing the new Duckiebots. In the final dataset, we have merged old and new Duckiebots to ensure that we can detect both variations.

Data Annotation Procedure

Modified 2021-10-31 by tanij

We used OpenCV’s free CVAT tool to annotate the dataset.