Generative AI for Perception
Using generative models to synthesize training data and bridge domain gaps for visual perception tasks.
Leveraging generative models — including diffusion models and image-to-image translation — to synthesize training data, augment scarce datasets, and improve model robustness across domains.
Key contributions:
- SSL-RGB2IR (IROS 2024) — Semi-supervised RGB-to-IR image translation for boosting segmentation and detection model training.
- Syn2Real (IEEE GRSL 2025) — Diffusion-based synthetic side-scan sonar data generation for underwater object detection, demonstrating significant improvements over real-data-only baselines.
- Generative fused RGB–IR and RGB–SAR visualizations for Sim-to-Real aerial imagery analysis.