A Greener AI-Based Crowd Counting via Efficient Deep Learning
Audrey Chrysler, Rivaldi Gunarso, Triladias Puteri, Tjeng Wawan Cenggoro
Last modified: 2021-06-14
Abstract
Similar to other implementations of Artificial Intelligence (AI), a crowd counting AI development tends to produce larger models over time to improve the performance. This trend negatively impacts the environment because the development of larger AI models generates more CO2 that causes global warming. This study instead focuses on the development of crowd counting AI that has a small size with competitive performance. Our proposed model achieved 11.6 MAE and 19.7 MSE on the ShanghaiTech Part B dataset, which meets the current standard in crowd counting research. Currently, our proposed model is also the 3rd best model in terms of performance-size tradeoff among all crowd counting AI with known model size.