sababishraq/foodsense-dataset

FoodSense Dataset

Human-Annnotated Sensory Ratings for Food Images

Human-annotated food images with sensory ratings (taste, smell, texture, sound) and free-text descriptors. This dataset was built to train models to reason about the cross-sensory properties of food just by looking at pictures.

Accepted to the CVPR 2026 Workshop on Meta Food.

Quick Start

You can load the dataset natively in Python using the datasets library. The viewer shows the metadata.csv which perfectly matches the images.

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("sababishraq/foodsense-dataset", split="train")

# View the first item
item = dataset[0]
print(f"Image File: {item['file_name']}")
print(f"Taste ({item['RescaledRating_taste']}/5): {item['taste_desc']}")
print(f"Smell ({item['RescaledRating_smell']}/5): {item['smell_desc']}")
print(f"Texture ({item['RescaledRating_texture']}/5): {item['texture_desc']}")
print(f"Sound ({item['RescaledRating_sound']}/5): {item['sound_desc']}")

# Display the image
item["image"].show()

Dataset summary

Statistic	Value
Total food images	2,987
Participants	8,382
Total annotations (rows)	66,842 participant–image pairs
Mean annotators per image	22.38 (SD 2.02)
Rescaled ratings (pipeline)	1–5 (RescaledRating_*)

Source images: Yelp Open Dataset.

Layout

• JPEGs at the repository root (ImageFolder convention).
• metadata.csv: Full table; file_name points to each JPEG for the Dataset Viewer. Image_Name is the study basename (what load_human_sensory_data matches unless file_name is used).

Sample images

	file_name	Preview
1	0001_01lamiW2bWW0rXlllNHYMA.jpg
2	0002_01zZeZBIFZ82S5XmA4GYJg.jpg
3	0003_05KUDlEPkMLF-fTrTk4qxQ.jpg

Columns

Column	Description
file_name	Canonical filename in this repo (Viewer).
participantId, Image_ID, Image_Number, Set_Number, Image_URL	Study metadata.
Image_Name	Basename as in the export (training loader).
taste…sound	Original 1–7 study ratings.
*_desc	Text descriptors.
sensory_average	Mean of four modalities (study scale).
CanInfer_*	Inferability flags (1 if human could infer this sensory dimension, 0 otherwise).
RescaledRating_*	1–5 targets for training.

Training code

Point --human_csv at metadata.csv and --image_dir at the folder containing the JPEGs (snapshot root). See dataset.py load_human_sensory_data.

Citation

@inproceedings{ishraq2026foodsense,
  title     = {FoodSense: A Multisensory Food Dataset and Benchmark for
               Predicting Taste, Smell, Texture, and Sound from Images},
  author    = {Ishraq, Sabab and Aarushi, Aarushi and Jiang, Juncai and Chen, Chen},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and
               Pattern Recognition (CVPR) Workshops},
  year      = {2026}
}

License: CC BY 4.0.