Source-linked AI summary
Does Object Recognition Work for Everyone?
Terrance DeVries, Ishan Misra, Changhan Wang, Laurens van der Maaten
TL;DR
The paper asks whether object-recognition systems work equally well for people across countries and income levels. Using the geographically diverse Dollar Street household-item dataset, it evaluates publicly available systems and finds substantial accuracy disparities, indicating that further work is needed.
Problem
Object-recognition systems should work equally well for users worldwide, but their performance across cultural and socio-economic contexts requires investigation.
Method
The study evaluates publicly available cloud-based object-recognition systems on Dollar Street photographs of household items, using country and purchasing-power-adjusted household-income metadata.
Results
Accuracy is approximately 10% lower for households earning less than US$50 per month than for households earning more than US$3,500 per month, with an absolute United States–Somalia or Burkina Faso difference of around 15−20%.
Takeaways & Limitations
The findings suggest that additional work is needed to develop object-recognition systems that work for people across countries and income levels.
Takeaways & Limitations
The study addresses representation bias and elements of measurement bias, but not historical, evaluation, or aggregation biases.
Abstract
from arXiv · showhide
The paper analyzes the accuracy of publicly available object-recognition systems on a geographically diverse dataset. This dataset contains household items and was designed to have a more representative geographical coverage than commonly used image datasets in object recognition. We find that the systems perform relatively poorly on household items that commonly occur in countries with a low household income. Qualitative analyses suggest the drop in performance is primarily due to appearance differences within an object class (e.g., dish soap) and due to items appearing in a different context (e.g., toothbrushes appearing outside of bathrooms). The results of our study suggest that further work is needed to make object-recognition systems work equally well for people across different countries and income levels.
1. Introduction
The paper examines whether object-recognition systems work equally well across countries and income levels, finding lower performance for household items common in non-Western and low-income communities. Accuracy disparities are consistent across commercial systems and suggest further work is needed.
- Object-recognition systems are deployed globally, creating a responsibility to work equally well regardless of cultural background or socio-economic status.
- Accuracy differs by around 15−20% between items photographed in the United States and those photographed in Somalia or Burkina Faso.
- These geographical and income-related disparities are consistent across a range of commercial cloud image-recognition services.
- The study uses household-item images spanning countries and income levels to illustrate classifications produced by publicly available recognition systems.
2. Measuring Recognition Performance of Household Items Across the World
The study evaluates publicly available object-recognition systems on the geographically and economically diverse Dollar Street household-item dataset. Across systems, recognition accuracy is substantially higher for high-income households and varies across countries.
- Dataset: The Dollar Street dataset contains 135 household-item classes photographed in 264 homes across 54 countries.
- Dataset: The analysis removes abstract classes and evaluates recognition accuracy on the remaining 117 classes.
- Experimental setup: Five cloud services plus a sixth object-recognition system are evaluated using location and purchasing-power-adjusted monthly household consumption income.
- Experimental setup: Human annotators assess accuracy@5 by checking whether any of the five highest-ranked predictions matches the dataset’s ground-truth class.
- Results: Across all six systems, household-item recognition accuracy is substantially higher for high-income households than for low-income households.
- Caveat: The study’s annotators were all based in the United States, which may bias annotations, although qualitative evaluations suggest the impact is very small.
- Results: Accuracy is approximately 15% higher for household items photographed in the United States than for those photographed in Somalia or Burkina Faso.
3. Sources of Accuracy Discrepancies
The paper attributes accuracy discrepancies to unrepresentative geographical sampling and English-centered data collection, with language-specific searches returning visually different images.
- Some Dollar Street labels reflect an object's affordance rather than its physical identity, such as “refrigerators” referring to cooling pots and pans.
- Computer-vision datasets undersample visual scenes in heavily populated regions, especially Africa, India, China, and South-East Asia.
- The same Flickr queries in Hindi and English return visually different results.
- English-centered image collection excludes images tagged in other languages and can introduce subtler class-coverage problems.
- The dataset distributions can be compared against a world population-density reference map to assess geographical representativeness.
4. Related Work
The work connects its empirical study to fairness and representative computer vision while emphasizing that multiclass image recognition does not fit many binary fairness formulations.
- The paper is related to broader research on fairness and representative computer-vision systems.
- Existing machine-learning fairness research includes formulations such as statistical parity and equalized odds, alongside impossibility results.
- This study focuses on multiclass and potentially multi-label prediction rather than binary prediction.
- Because the image input lacks user or group-membership information, group-based fairness formulations are difficult to apply directly.
- Related computer-vision studies have documented disparities in gender classification, image-search representation, and ImageNet class associations.
5. Discussion
The study identifies geographical and income-related accuracy disparities but addresses only representation and elements of measurement bias, leaving other bias sources unresolved.
- The analysis addresses representation bias and elements of measurement bias, but not historical, evaluation, or aggregation bias.
- The study identifies geographical and income-related accuracy disparities without solving them.
- Suggested mitigation approaches include geography-based resampling and multilingual training, but they may be insufficient on their own.
- The authors suggest that broader solutions may require few-shot learning of new visual classes and reduced sensitivity to training-data variation.
Supplemental Material
The supplemental material documents annotation procedures, dataset classes, country and income analyses, and visualizations of accuracy disparities across income levels and countries.
- Annotation: Human annotators assessed whether any of five top-ranked predictions matched each image's ground-truth class annotation.
- Dataset: The analysis used 117 Dollar Street classes comprising 20,455 total images.
- Income disparities: Among the ten classes with the largest income-group accuracy discrepancies, some labels reflect affordances rather than the depicted object.
- Income disparities: In India, accuracy across 2,221 images remains correlated with household income after holding country constant, suggesting location and income both partly drive the observations.
- Income disparities: The supplemental figures include the ten classes with the largest differences between the highest- and lowest-income groups.
- Geographical disparities: A country-level Facebook-system map encodes approximately 60%, 75%, and 90% accuracy with red, yellow, and green, respectively.