SynBlink and BlinkFormer: A Synthetic Dataset and Transformer-Based Method for Video Blink Detection

Accurate blink detection algorithms have significant implications in numerous fields, including human-computer interaction, driving safety, cognitive science, and medical diagnostics. Despite considerable efforts, the dataset volume for blink detection remains relatively small due to the cost of data collection and annotation, and there is still room for improvement in the accuracy of current algorithms. In this paper, we introduce a workflow for synthesizing video data in Blender. Fully-rigged 3D human models are programmatically controlled, with variations in head movement, blinking, camera angles, background types, and lighting intensities. We used this workflow to create the SynBlink dataset, which includes 50,000 video clips and their corresponding annotations. Additionally, we present BlinkFormer, an innovative blink detection algorithm based on Transformer architecture that fully exploits temporal information from video clips. The model not only detects blinks for the entire input video but also estimates blink strength for each frame individually. Experimental results reveal that the BlinkFormer outperforms other state-of-the-art blink detection methods, achieving the highest F1-score on HUST-LEBW dataset. This accomplishment highlights the effectiveness of our approach in accurately detecting blinks and its potential for real-world applications.