CHOWNET

The challenge

Food is deeply connected to culture, identity, and everyday life. Yet, many computer vision systems still perform poorly when asked to recognise African foods, especially local dishes that are under-represented in mainstream image datasets.

This creates a wider problem for AI inclusion. If the foods, objects, and cultural realities of African communities are missing from datasets, then AI systems will continue to underperform for African users and African contexts.

For Nigerian cuisine in particular, many local dishes are visually diverse, served in different forms, and often contain multiple food items in one image. This makes food recognition more challenging and highlights the need for carefully curated, locally grounded datasets.

Our approach

CHOWNET takes a community-driven approach to building visual datasets for African food. The project began as an AI Saturdays Lagos community initiative to collect images of foods commonly eaten in Nigeria, with the long-term goal of improving how AI systems understand local African cuisine.

Data

CHOWNET-V1 contains 118 human-annotated food images with 99 unique labels. The dataset was curated from an initial community collection of 568 image submissions across 15 food categories.

The dataset was cleaned carefully to improve quality and ensure responsible release. This included removing duplicates and near-duplicates, excluding blurry images, removing images with privacy concerns, and checking for copyright issues.

CHOWNET-V1 is available for download on Zenodo.

Tasks

CHOWNET-V1 supports several computer vision research tasks, including:

• Food clustering
• Multi-label classification
• Food object detection
• Food captioning

These tasks make the dataset useful for researchers working on food recognition, cultural representation in AI, and computer vision for African contexts.

Benchmarks

CHOWNET provides a foundation for evaluating how well computer vision models recognise Nigerian food items. By creating a clean, annotated dataset of local food images, the project supports more representative benchmarking and encourages the broader AI community to test models beyond common Western food categories.

Get involved

We welcome researchers, computer vision practitioners, data annotators, food communities, and contributors interested in improving African representation in AI datasets.

You can contribute by supporting future data collection, helping annotate food images, running experiments, or collaborating on research around African food recognition.

Reach out at research@tri-ai.org

Dataset

CHOWNET-V1: An Image Dataset of Nigerian Food

A human-annotated dataset of Nigerian food images for multi-label classification, object detection, clustering, and captioning.

Available on Zenodo: https://zenodo.org/records/13633554

Poster presentation

CHOWNET: An Image Dataset for Local African Food

Authors: Tejumade Afonja, Oluwafemi Azeez, George Igwegbe
https://instadeep.com/2018/12/instadeep-presents-two-ai-research-papers-at-neurips-2018/