B&Q Research Dataset

Brooklyn & Queens

Near/Remote Sensing

A large-scale dataset for urban semantic segmentation, featuring paired high-resolution overhead and ground-level street-view imagery.

Overhead Images

53,000+

Panoramas

175,000+

Tasks

5

Citations

100+

Brooklyn & Queens
Download Dataset

Reference Publications

Papers that introduced this dataset

Revisiting Near/Remote Sensing with Geospatial Attention

Scott Workman, M. Usman Rafique, Hunter Blanton, and Nathan Jacobs

in IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

A Unified Model for Near and Remote Sensing

Scott Workman, Menghua Zhai, David J. Crandall, and Nathan Jacobs

in IEEE International Conference on Computer Vision (ICCV)

Dataset Details

Technical specifications

The Brooklyn & Queens (B&Q) dataset was established to support the task of near/remote sensing—the joint inference of geospatial functions from co-located overhead and ground-level imagery. By combining the dense coverage of high-resolution overhead imagery with the semantically rich details of street-level panoramas, the dataset enables the creation of high-precision maps for urban properties that are often ambiguous from a single perspective. Originally introduced for urban semantic segmentation, the dataset was later extended to include additional per-pixel labeling tasks, including physical metrics derived from LiDAR, providing a city-scale benchmark for multimodal urban analysis.

Dataset Features

Remote Sensing

  • GSD: ~30 cm ground sample distance (Bing Maps Zoom 19).
  • Coverage: 53,000+ dense overhead image tiles across NYC.

Proximate Sensing

  • Format: Full 360° Equirectangular Street View panoramas.
  • Scale: 175,000+ panoramas georeferenced to overhead tiles.

LiDAR Extension

  • Source: 2017 topographic/bathymetric LiDAR capture.
  • Metrics: Derived per-pixel height and 6-inch land-cover.

Joint Modeling

  • Scope: Supporting research into joint inference over co-located image pairs.
  • Labels: 5 per-pixel tasks spanning semantic and physical categories.

The B&Q dataset features five labeling tasks that evaluate the impact of combining remote and proximate imagery for urban analysis.

Land Use

Classification of urban functional parcels (e.g., residential).

Building Function

Identifying societal purpose (e.g., educational, religious).

Building Age

Estimation of construction periods from municipal records.

Land Cover

8-class segmentation derived from 6-inch LiDAR data.

Building Height

Metric vertical estimation using Highest-Hit DSMs.