Bridging the Gap Between Events and Frames Through Unsupervised Domain Adaptation

Reliable perception during fast motion maneuvers or in high dynamic range environments is crucial for robotic systems. Since event cameras are robust to these challenging conditions, they have great potential increase the reliability of robot vision. However, event-based vision has been held back by shortage labeled datasets due novelty cameras. To overcome this drawback, we propose a task transfer method train models directly with images and unlabeled data. Compared previous approaches, (i) our transfers from single events instead frame rate videos, (ii) does not rely on paired sensor achieve this, leverage generative model split features into content features. This enables efficient matching between latent spaces images, which successful transfer. Thus, approach unlocks vast amount existing image training neural networks. Our consistently outperforms methods targeting Unsupervised Domain Adaptation object detection 0.26 mAP (increase 93%) classification 2.7% accuracy.