Gaze estimation optimization through performance analysis of ResNet variants

The performance of different ResNet variations for the gaze estimation task is evaluated in this research. Although many studies utilize various ResNet variants as backbone for gaze estimation, the selection of a specific ResNet variant is often made without a detailed justification, despite the differences in their architecture, depth, and computational complexity. To address this gap and optimise the gaze estimation algorithm guiding model selection and find the best trade-off between model complexity and gaze estimation accuracy, a thorough analysis is carried out. Eth Gaze data collected under varying conditions has been used for model training and validation. The research evaluates six ResNet variants which include ResNet18, ResNet34, ResNet50, ResNet101, ResNet152, and ResNext50_32x4d. Angular error, mean, mean absolute, and root mean square errors (MSE, MAE, RMSE), R2 score, and computing efficiency based on model size, FLOPs, inference time, are among the evaluation criteria. Higher accuracy can be achieved with deeper ResNet models, but the cost of computing is substantial, as verified by this research. Notably, among other findings, ResNext50_32x4d emerges as an optimal choice, offering an exceptional balance between accuracy and computational efficiency for gaze estimation tasks. This variant demonstrates superior performance with lower computational demands, making it particularly suitable for real-time gaze estimation applications where resources may be limited. This works findings provide valuable insights into the trade-offs between model complexity and performance, guiding the selection of appropriate architectures for practical implementations of gaze estimation for various applications like human-computer interaction, driver attention, neuromarketing and more.