ESPA-Face: A Framework for Event Self-Supervised Pretraining and Augmentation for Face Detection

초록

Event cameras operate asynchronously, capturing pixel brightness changes, but manual labeling of such data is both time-consuming and expensive, with limited dataset availability. To reduce dependency on labeled event data, we propose ESPA-Face, a two-stage learning framework for event camera-based face detection. Initially, self-supervised learning is applied to real event data without annotations to pretrain a model. This pretrained model is then used as the backbone for the downstream face detection task, which is fine-tuned using a small, annotated synthetic event dataset. Additionally, we introduce an event augmentation method tailored to reflect the asynchronous nature of event data, enhancing detection accuracy. Our comprehensive evaluation, conducted on both the public N-Caltech101 and self-collected event datasets with diverse motions, demonstrates the robustness and generalizability of ESPA-Face. The proposed method achieves a precision of 99.77% on the N-Caltech101 dataset and 97.30% on our self-collected dataset, demonstrating strong performance on both real-event evaluation datasets.

키워드