cvlab/new-york-smells

New York Smells: A Large Multimodal Dataset for Olfaction

While olfaction is central to how animals perceive the world, this rich chemical sensory modality remains largely inaccessible to machines. One key bottleneck is the lack of diverse, multimodal olfactory data collected in natural settings. We present New York Smells, a large-scale dataset of paired image and olfactory signals captured in-the-wild. Our dataset contains 7,000 smell-image pairs from 3,500 distinct objects across diverse indoor and outdoor environments, and it is 70x larger than prior olfactory datasets.

• Project page: https://smell.cs.columbia.edu
• Paper: https://arxiv.org/abs/2511.20544

Quickstart

from datasets import load_dataset

# Sensors, labels and metadata only -- 44 MB, two files.
ds = load_dataset("cvlab/new-york-smells", "olfaction", split="test")

# Everything, including the full-resolution image.
ds = load_dataset("cvlab/new-york-smells", split="train")

row = ds[0]
row["image"]              # PIL.Image, 1920x1080, lossless
row["sample_raw"]         # (T, 32) sensor response during the sample
row["baseline_raw"]       # (T, 32) sensor baseline before it
row["smellprint_vector"]  # (32,) hand-crafted feature
row["material"]           # class index; ds.features["material"].int2str(...) for the name

One row is one complete sample -- no joins between files.

Config	Contents	Size	Files
olfaction	sensors, metadata, labels, splits	44 MB	2
default	the above plus the full-resolution PNG	16 GB	~35

The olfaction config is enough to reproduce the scene / material / object recognition results without downloading any images.

train / test follow the paper's object-level split (5,868 / 1,036); the split_uniform and split_session columns carry the other two definitions. The parquet rows are a superset of metadata.jsonl -- every column below is present under the same name, plus image, sample_raw, baseline_raw, and scene / object / material as integer class labels.

One deliberate difference: smellprint_vector in the parquet is read from smellprint_npy/*.npy at full float64 precision, whereas the copy in metadata.jsonl was rounded to about five significant digits by JSON serialisation. Prefer the parquet (or the .npy) when the smellprint is a model input.

The per-session file tree below remains in this repository and is still the only source for depth, depth_raw, rs_rgb and clip_features. Use hf download --include "<session>/depth/*" to fetch those selectively.

Layout

Each top-level 2025-MM-DD_HH-MM-SS directory is one recording session, and the files within it correspond to samples from that session.

2025-MM-DD_HH-MM-SS/
├── baseline_raw        Olfactory sample for baseline (ambient) smell
├── sample_raw          Olfactory sample for the object
├── clip_features       Pre-computed CLIP features
├── depth               Depth camera visualization
├── depth_raw           Raw depth camera values
├── rgb                 RGB image from the iPhone
├── rs_rgb              RGB image from the RealSense
├── sample_metadata     Meta-data about the sample
└── smellprint          Handcrafted smellprint feature (baseline)

Labels

Object and material labels from GPT-4o are included at the top level:

materials.json             Material taxonomy
material_labels.json       Per-sample material labels
object_labels/
├── object_labels.json     Per-sample object labels
├── object_clusters.json   Object label clusters
└── label_to_cluster.json  Object label -> cluster name

material_labels.json and object_labels/object_labels.json are both keyed by session, then by sample index within that session:

{"2025-MM-DD_HH-MM-SS": {"0": { ... }, "1": { ... }}}

Each entry lists the images the label was derived from, as paths relative to the root of this repository (rgb_paths, rs_rgb_paths). Material entries additionally carry label, form, composition and state; object entries carry label and label_index.

materials.json maps each material name to its {form, composition, state} taxonomy. Note that its values are capitalized (Ferrous, Metal, Abiotic) whereas the per-sample values in material_labels.json are lowercase — compare case-insensitively.

object_clusters.json groups the 146 object labels into 49 coarser clusters, and label_to_cluster.json is the flat label-to-cluster mapping.

Scene labels are per-session rather than per-sample:

setting_labels.json        Session -> one of 8 scene categories, plus free-text location
indoor_labels.json         Session -> indoors / outdoors
material_to_idx.txt        Frozen material name -> class index (51 classes)

material_to_idx.txt fixes the class ordering used by the paper's material classifier, so heads trained at different times stay comparable. The 8 scene categories are campus building, campus outdoors, apartment, street/park, dining hall, library, gym, office.

Note: indoor_labels.json labels session 2025-04-12_16-59-24 ("CV Lab Lounge") as outdoors while setting_labels.json calls it a campus building. This is the only disagreement between the two files and appears to be a labelling error; it is shipped as-is because it is what the published experiments used.

Label coverage: 3,452 samples across all 60 sessions, with 146 distinct object labels and 54 distinct material labels.

Splits

splits.json contains three train/test splits, all covering the 6,904 samples from the 58 sessions with n_samples > 0 (the 30 unlabeled samples are excluded).

The id order in splits.json is significant. The paper's N = 933 retrieval gallery is a positional random_split(seed=42) over the test list, so re-sorting the ids selects a different gallery. Preserve the order as shipped.

objectlevel — the split used in the paper. Object ids are sampled uniformly at random 85/15, and both samples of an object follow it, so an object's two samples never straddle the train/test boundary.

	train	test
samples	5,868	1,036
objects	2,934	518

uniform — sampled uniformly at random 85/15 over samples rather than objects, so 844 objects have one sample in train and its twin in test. Provided because the paper's recognition results (Table 2) used it; prefer objectlevel for new work, and read numbers on uniform as an upper bound.

	train	test
samples	5,868	1,036
objects with samples on both sides	—	844

session — grouped by capture day, so sessions recorded on the same day never straddle the boundary: 52 / 6 sessions (6,134 / 770 samples) across 19 / 3 days.

All three are also available as the split_objectlevel, split_uniform and split_session columns in metadata.jsonl, so the viewer can be filtered by split directly.

metadata.jsonl

metadata.jsonl at the repository root has one row per sample (6,934 rows) and drives the dataset viewer. It references the media files by relative path rather than duplicating them, so the full-resolution originals are the ones shown.

Column	Description
rgb_file_name	RGB image from the iPhone
rs_rgb_file_name	RGB image from the RealSense
smellprint_file_name	Smellprint visualization
smellprint_vector	Handcrafted smellprint, 32 floats (one per Cyranose sensor)
session, location	Recording session id and free-text capture location
object_idx, sample_idx, global_id	Identifiers; the two samples of one object share object_idx
object_label	GPT-4o object label
material_label, material_form, material_composition, material_state	GPT-4o material labels
temperature, humidity, pid	Sensor readings at capture time
setting, indoor	Per-session scene category and indoors/outdoors
split_objectlevel, split_uniform, split_session	train / test per the three splits below

Depth (depth, depth_raw), raw olfactory measurements (sample_raw, baseline_raw, each timesteps x 32 sensors) and clip_features are present in the repository but are not referenced from metadata.jsonl.

30 rows (the 15 objects without labels, x2 samples) have nulls for the label, rs_rgb_file_name and smellprint_file_name fields.

License

Released under CC BY 4.0. See LICENSE for the full terms.

Citation

@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}
}