neurocat's Augmented-BDD dataset

The Augmented-BDD (A-BDD) dataset consists of 63,735 augmented images, alongside the original 1,820 non-augmened source images with associated labels (e.g. segmentation and bounding boxes) and experiment replication data.

A-BDD's augmentations fall into seven catagories emulating different weather conditions: overcast, sunglare and shadow, fog, and four types of rain (puddles, wet streets and droplets, streaks, and a composite). All catatgories has five intensity levels, resulting in 35 total subsets.

The A-BDD draws its source images from the Berkeley DeepDrive (BDD) 100K dataset. Below you can find also our research paper on the data and further information on the dataset and how it was produced. 

Research

The A-BDD white paper: What will you research?

Alongside the dataset our team has written "A-BDD: Diverse Adverse Weather and Lighting Conditions through Data Augmentations" (its replication data is included in the dataset).

In addition to presenting the A-BDD, the paper oulines the results of experiments to prove the validity (e.g. realism) of augmented data, evidence how it can improve ML (re)training, and showcase novel techniques to identify promising synthetic data for a given use case.

We believe the paper demonstrates that augmented data can play a pivotal role in closing performance gaps in adverse weather conditions. But don't take our word for it: read the paper, replicate our research, and conduct your own experiments with the dataset.

aidkit: Our tool for data augmentation

About the dataset

The A-BDD dataset was created using neurocat's proprietary aidkit software. This software takes a source image and, based on selected parameters, scalably generates variations of that image with the specified condition (e.g. heavy rain with reflective streets). To see what we mean, try out the web version of this augmentation tool on our Playground.

Our purpose in making this dataset public is to encourage research into the realism, applications, and other aspects of augmented data. In this spirit, if your research interest requires further augmentations, do not hesitate to reach out to us.

The A-BDD dataset uses images from the Berkeley DeepDrive (BDD) 100K dataset. As a modified version of the BDD100K, it is subject to the license available here. 

Driving safe perception

Get in touch

Interested? You can get the dataset now on Zenodo. If you want to learn more about the dataset, or the tool with which it was made, feel free to get in touch with us!

© neurocat GmbH