Voxel51/NYC_Smells

Dataset Card for New York Smells

[image]

This is a FiftyOne dataset with 20000 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/NYC_Smells")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

New York Smells is a large-scale multimodal dataset pairing visual imagery with electronic nose (e-nose) olfactory signals captured "in the wild" throughout New York City. The dataset enables cross-modal learning between smell and sight, addressing a critical gap in machine perception research where olfaction remains largely inaccessible to machines compared to vision, sound, and touch.

The dataset contains 7,000 smell-image pairs from 3,500 distinct objects across 60 recording sessions in diverse indoor and outdoor environments—approximately 70× more objects than existing olfactory datasets.

• —Curated by: Ege Ozguroglu, Junbang Liang, Ruoshi Liu, Mia Chiquier, Michael DeTienne, Wesley Wei Qian, Alexandra Horowitz, Andrew Owens, Carl Vondrick
• —Institutions: Columbia University, Cornell University, Osmo Labs
• —Funded by: Not specified
• —License: Not specified (check project website for updates)

Dataset Sources

• —Repository: https://smell.cs.columbia.edu/
• —Paper: https://arxiv.org/abs/2511.20544
• —Data Download: https://smell.cs.columbia.edu/static/smell-dataset.tar.gz (27 GB)
• —Hardware Rig Specs: https://smell.cs.columbia.edu/static/hardware.zip

Uses

Direct Use

• —Cross-modal smell-to-image retrieval: Given a query smell, retrieve matching images in embedding space
• —Scene/object/material recognition from smell alone: Classify scenes, objects, and materials using only olfactory signals
• —Fine-grained olfactory discrimination: Distinguish between similar objects (e.g., different grass species)
• —Olfactory representation learning: Train general-purpose smell embeddings using visual supervision
• —Multimodal sensor fusion research: Combine RGB, depth, and chemical sensor modalities

Out-of-Scope Use

• —Medical diagnosis or health-related smell detection (dataset not collected for clinical purposes)
• —Hazardous material detection (not designed for safety-critical applications)
• —Individual identification or tracking via smell
• —Production deployment without additional validation on target domain

Dataset Structure

FiftyOne Dataset Organization

The dataset is loaded as a grouped dataset with three slices per sample:

Sample Fields

Pre-computed Visualizations (Brain Keys)

Dataset Creation

Curation Rationale

While olfaction is central to how animals perceive the world, this sensory modality remains largely inaccessible to machines. A key bottleneck is the lack of diverse, multimodal olfactory training data collected in natural settings. Existing olfactory datasets are small and captured in controlled lab environments. New York Smells addresses this gap by providing large-scale, in-the-wild paired vision-olfaction data.

Source Data

Data Collection and Processing

• —Sensor Hardware: Cyranose 320 electronic nose with 32 polymer-composite chemoresistive sensors, mounted on a custom 3D-printed rig with an iPhone camera and Intel RealSense depth sensor
• —Collection Method: Researchers walked through various NYC locations capturing synchronized images and e-nose readings of odorant objects
• —Locations: Parks, gyms, dining halls, libraries, streets, and other indoor/outdoor environments across New York City
• —Sessions: 60 recording sessions
• —Samples: 7,000 smell-image pairs from 3,500 distinct objects
• —Additional Sensors: RGB-D, temperature, humidity, volatile organic compound (VOC) concentration

Olfactory Signal Format

• —Baseline: ~17 timesteps × 32 sensors — ambient air reading before approaching object
• —Sample: ~17 timesteps × 32 sensors — reading while near object
• —Smellprint: 32-element vector — pre-computed time-collapsed fingerprint
• —Values: Normalized resistance ratios (ΔR/R₀)

Who are the source data producers?

Research team from Columbia University, Cornell University, and Osmo Labs collected all data in New York City.

Annotations

Annotation process

• —Location labels: Manually recorded during collection sessions
• —Object indices: Assigned during collection (human-readable labels pending release)
• —Scene/object/material labels: Generated via GPT-4o (release pending per authors)
• —CLIP features: Pre-computed using CLIP model on RGB images

Who are the annotators?

The research team annotated location and object information. GPT-4o was used for scene/object/material labeling (pending release).

Citation

BibTeX

@article{ozguroglu2025smell,
  title={New York Smells: A Large Multimodal Dataset for Olfaction},
  author={Ozguroglu, Ege and Liang, Junbang and Liu, Ruoshi and Chiquier, Mia and DeTienne, Michael and Qian, Wesley Wei and Horowitz, Alexandra and Owens, Andrew and Vondrick, Carl},
  journal={arXiv preprint arXiv:2511.20544},
  year={2025}
}

APA

Ozguroglu, E., Liang, J., Liu, R., Chiquier, M., DeTienne, M., Qian, W. W., Horowitz, A., Owens, A., & Vondrick, C. (2025). New York Smells: A Large Multimodal Dataset for Olfaction. arXiv preprint arXiv:2511.20544.