preloader

A Diverse Driving Dataset for Heterogeneous Multitask Learning

Paper Code Doc Data Discuss


Explore it

Main Features

Large-scale

100K driving videos (40 seconds each) collected from more than 50K rides, covering New York, San Francisco Bay Area, and other regions.

Diverse

Contains diverse scene types such as city streets, residential areas, and highways. Recorded in diverse weather conditions at different times of the day.

Multi-task

10 taksk in total: Lane detection, object detection, semantic segmentation, instance segmentation, multi-object tracking, segmentation tracking and more.

720p

High resolution

30fps

High frame rate

GPS/IMU

Trajectories

50k rides

Crowd sourced

Multiple Tasks

Object Detection

70,000/10,000/20,000 images for train/val/test, 1.8M objects.

Instance Segmentation

7,000/1,000/2,000 images for train/val/test, 120K masks.

Multi Object Tracking

1,400/200/400 videos for train/val/test, 160K instances, 4M objects.

Segmentation Tracking

154/32/37 videos for train/val/test, 25K instances, 480K masks.

Semantic Segmentation

7,000/1,000/2,000 images for train/val/test, 40 object classes.

Lane Marking

70,000/10,000/20,000 images for train/val/test, 8 main categories.

Drivable Area

70,000/10,000/20,000 images for train/val/test, 8 main categories.

Image Tagging

6 weather conditions, 6 scene types, 3 distinct times of the day

Imitation Learning

GPS/IMU recordings with visual input and the driving trajectories.

Domain Adaptation

Diverse weather, road and daytime conditions.

Docs & Tools

We provide documents and tools for inspection, preparation, and evaluation of the BDD100K dataset.

Data Download

You can simply log in and download the data in your browser after agreeing to BDD100K license.

Visualization

We provides scripts to parse and visualize the labels, and a tool to display the trajectories.

Label Format

We use a consistent data annotation format across all different tasks. We choose the Scalabel [link] format for this.

Evaluation

We provide evaluation scripts, online testing servers and challenges to verify your algorithm.

Leaderboard

Object Detection

open end

  • 14,000/2,000/4,000 images
  • 1.8M objects
Participate

Instance Segmentation

open end

  • 7,000/1,000/2,000 images
  • 120K objects
Participate

BDD100K

Facilitate algorithmic study on large-scale diverse visual data and multiple tasks

Download

Label Visualization

Object Detection

Object Detection

Object Detection

Object Detection

Instance Segmentation

Instance Segmentation

Instance Segmentation

Instance Segmentation

Instance Segmentation

Instance Segmentation

Box Tracking

Box Tracking

Box Tracking

Box Tracking

Box Tracking

Box Tracking

Seg Tracking

Seg Tracking

Seg Tracking

Seg Tracking

Seg Tracking

Seg Tracking

Scalabel

Scalabel

BDD100K is compatible with the labels generated by Scalabel. The labels are released in Scalabel Format.