DisDop: Distillation with Domain Priors for Open-Vocabulary Aerial Object Detection
Researchers have developed DisDop, a new framework designed to improve open-vocabulary object detection in aerial imagery. This method systematically distills domain-specific knowledge from foundation models like RemoteCLIP and DINOv3 into a more lightweight detector. DisDop enhances detection by combining visual priors from fused teacher models and textual priors from RemoteCLIP's text encoder, while also incorporating global context to better identify small objects. The framework has demonstrated state-of-the-art performance on relevant benchmarks. AI
IMPACT Improves accuracy for object detection in drone imagery, potentially benefiting applications like surveillance and mapping.