Berkeley Deep Drive Semantic Segmentation (BDD100K)

Attribute	Value
pretty_name	BDD100K Dataset
annotations_creators
language_creators
languages
licenses	BSD 3-Clause License
multilinguality
size_categories	1K<n<10K
source_datasets
task_categories	semantic-segmentation
task_ids
paperswithcode_id	bdd100k

Dataset Description

• Homepage: Berkeley Deep Drive Semantic Segmentation (BDD100K) dataset

• Licenses: BSD 3-Clause License

Dataset Summary

BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning

Download and prepare data

Login or register at this website, download the segmentation and 10K Images parts and extract them. Make sure to copy both the images folder as well as the labels folder in the same directory and use that below. Replace {PATH_TO_DATA} below with the location of the BDD100k folder. Use the following code to load it:

from squirrel_datasets_core.datasets.bdd100k import BDD100KDriver
iter_train = BDD100KDriver("{PATH_TO_DATA}").get_iter("train")
iter_val = BDD100KDriver("{PATH_TO_DATA}").get_iter("val")
iter_test = BDD100KDriver("{PATH_TO_DATA}").get_iter("test")

Dataset Structure

Data Instances

A sample from the training set is provided below:

{
    'image_url': '{PATH_TO_DATA}/train/0016E5_07920.png',
    'label_url': '{PATH_TO_DATA}/trainannot/0016E5_07920.png',
    'split': 'train'
    'image': array(...)
    'label': array(...)
}