View-Agnostic Domain Generalization
Segmentation and detection models that generalize across aerial and ground viewpoints without view-specific retraining.
Developing perception models that remain robust across drastically different viewpoints — drone, aerial, and ground-level — without requiring view-specific adaptation or retraining.
Key contributions:
- AetherVision-Bench (CVPRW 2025) — Large open-vocabulary RGB-IR benchmark covering multi-angle segmentation across aerial and ground perspectives.
- Pixel2Perspective (arXiv preprint, to appear) — Expands AetherVision-Bench to 13 datasets, adds an object-localization track for foundation VLMs, and introduces MSRIQ for quantifying cross-modal RGB-IR segmentation uncertainty. Project page
- SAGA (CVPRW 2025) — Semantic-aware gray color augmentation for visible-to-thermal cross-modal adaptation across drone and ground-based vision systems.
- VISTA-CLIP (CVPRW 2025) — Visual incremental self-tuned adaptation for continual panoptic segmentation across diverse domain shifts.