Scope
In (Fu et al, 2021) the authors investigate the carbon footprint of Computer Vision methodologies by decomposing the total CO2 emissions into two components: the emissions incurred during the initial model development phase and those accumulated during the model’s lifetime evaluation (training and usage). The paper provides quantitative metrics based on GPU hours, FLOPS, and power usage, and it discusses the ethical implications of high CO2 outputs in Computer Vision research area.
Summary
This study analyses CO2 emissions from Computer Vision be examining both the model-development phase and the evaluation phase. It finds that while initial phase has a significant CO2 cost, the evaluation phase dominated the overall emissions. The authors propose the integration of CO2 considerations into model design and evaluation can curb these emissions and align ethical AI practices with environmental sustainability.
Relevance for EXIGENCE
Since this work quantifies the environmental impact of Computer Vision model and suggests optimisation strategies, it provides valuable insights and methodologies that can be adapted within EXIGENCE to develop incentive mechanisms and improve energy efficiency in AI-driven systems.
- Fu, M. S. Hosseini, and K. N. Plataniotis, “Reconsidering CO2 emissions from Computer Vision“, IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 2311-2317. 2021.
