MAFM-Gaze: Multi-scale adaptive feature modulation for in-vehicle gaze estimation

Gaze estimation is a critical component of Driver Monitoring and Assistance Systems (DMAS), as it effectively identifies driver distraction and fatigue during driving, thus enhancing driving safety. Existing methods face challenges in achieving accurate gaze estimation in driving scenarios, due to complex factors such as illumination variation, facial occlusion, and extreme head poses. Therefore, an in-vehicle gaze estimation method with multi-scale adaptive feature modulation (MAFM-Gaze) is proposed in this paper. In MAFM-Gaze, the model takes only facial images as input and employs a pruned cross-stage partial (CSP) network to extract multi-scale features efficiently. A 3D Spatially Adaptive Feature Modulation (3D-SAFM) module is integrated into the feature mixing network, incorporating a multi-head concept with independent computation to fully exploit multi-scale features at each level, thereby enriching the current layer with critical global information and long-range dependencies. Additionally, a 3D Multi-scale Feature Fusion Module (3D-MFFM) is introduced to extract scale-invariant information and capture deeper interrelationships among multi-scale features. Experimental results shown that our model outperforms existing state-of-the-art in-vehicle gaze estimation methods, achieving a mean angular error of 6.65°with a compact model size of only 44.6MB. The code will be available at https://github.com/zhanggan123456/MAFM-Gaze.