Efficient event-based object detection with lightweight CNNs: a surprising advantage over transformers

Achieving robust robotic perception in highly dynamic environments requires accurate and low-latency visual object detection. However, conventional RGB cameras often exhibit limited performance in such fast-changing scenarios due to significant blind time, imaging latency, and motion blur. In this context, event cameras have emerged as a promising solution, offering microsecond-level temporal resolution, high dynamic range, and immunity to motion blur. However, despite these significant theoretical advantages, algorithms for event cameras are still at a nascent stage. Existing approaches, including both transformers and traditional CNNs, struggle to achieve high accuracy and low latency simultaneously, thereby limiting their effectiveness for object detection in highly dynamic environments. Notably, recent studies such as ConvNeXt suggest that modernized CNNs can outperform both vision transformers and traditional CNNs by adopting transformer-like architectural designs. Motivated by this insight, we propose a lightweight CNN-based object detection model, EvConvLite. By strategically "modernizing" the traditional CNNs, our model not only achieves 3× inference speedup, but also surpasses state-of-the-art transformer-based models in detection accuracy. To the best of our knowledge, this is the first reported instance of a lightweight CNN-based model outperforming transformers in event camera object detection accuracy.