YOLO-World for Aerial Zero-Shot Detection Through Alignment and Bottleneck Analysis

Abstract

Open-vocabulary object detectors have achieved strong zero-shot performance on natural-image benchmarks, but they collapse to near-zero novel-class mAP on aerial imagery. This study focuses on diagnosing this failure on the DIOR 16/4 generalized zero-shot detection benchmark through a sequence of controlled probes on YOLO-World. Guided by these diagnoses, a simple model also be designed to validate - RemoteCLIP align YOLO-World (RC-YOLO World). The resulting model improves novel@50 from $\mathbf{1. 8 2 \%}$ to $\mathbf{7. 8 1 \%}$ a gain over the strongest textencoder-swap baseline without pseudo-labeling, auxiliary heads, or additional detection data. The value of the analysis is not the remedy itself, but the diagnostic framework it validates, which should be routine checks before deploying open-vocabulary detectors in specialized domains.