Two-Level Cross-Modal Domain Adaptation for Event-Based Obstacle Detection via Asymmetric Pixel Translation and Contrastive Feature Alignment

Event-based cameras have emerged as promising sensors for autonomous systems due to their high temporal resolution, low latency, and resilience to challenging illumination. However, deploying event cameras for obstacle detection requires large annotated datasets that are prohibitively expensive to collect, limiting the practical adoption of event-based perception in autonomous systems. To overcome this bottleneck, we propose an unsupervised two-level domain adaptation (UDA) framework that transfers obstacle-detection knowledge from abundantly labeled RGB imagery to unlabeled event streams, eliminating the need for manual annotation in the event domain. At the first level, a pixel-level translation module employs asymmetric bidirectional generators with event-domain adversarial supervision between synthetic RGB and real event domains. At the second level, dual encoders extract semantically aligned representations enhanced through hierarchical and sparse attention mechanisms that capture dense spatial structures and motion-salient cues. The unified detection head, built upon a compact YOLO architecture and inspired by the obstacle-detection branch of J-MOD${}^2$, jointly predicts bounding boxes, objectness confidence, class identity, and mean obstacle distance. The network is trained on labeled source data and fine-tuned on unlabeled target data using confidence-based pseudo-labels, enabling label-free adaptation. Extensive experiments on four event-based benchmarks Gen1, 1Mpx, DSEC, and MVSEC demonstrate that the proposed framework achieves competitive mAP@0.5:0.95 scores of 0.444, 0.438, 0.465, and 0.474, respectively. On MVSEC, the framework outperforms the closest image-to-event UDA baseline under a comparable car-detection protocol, while maintaining low latency and compact model complexity. To the best of our knowledge, this work represents one of the first comprehensive UDA frameworks for obstacle detection in event-based vision, establishing a new direction toward efficient, domain-invariant perception under asynchronous sensing.