Reconsidering CO2 Emissions From Computer Vision

Abstract

The Computer Vision community has achieved remarkable progress, but this progress comes with an environmental cost. This paper examines the CO2 emissions associated with training CV models and proposes strategies for reducing the community's carbon footprint.

Key Contributions

1. Emissions Analysis: We quantify the carbon emissions associated with training popular CV models.

1. Impact Assessment: We assess the cumulative environmental impact of the CV research community.

1. Mitigation Strategies: We propose practical strategies for reducing emissions while maintaining research progress.

The Carbon Cost of CV

• Training large models can emit as much CO2 as multiple transatlantic flights
• The cumulative impact of thousands of experiments compounds significantly
• Hardware efficiency improvements alone cannot offset growing model sizes

Proposed Solutions

• Efficient architecture design
• Transfer learning and model reuse
• Carbon-aware scheduling of experiments
• Reporting emissions alongside accuracy metrics

Community Responsibility

As AI systems become more prevalent, the CV community has a responsibility to lead by example in sustainable research practices. This includes considering environmental impact as a first-class metric alongside accuracy and efficiency.

Call to Action

We encourage conferences to require carbon reporting and for researchers to prioritize efficiency alongside performance.