Automated Cross-Domain Generalization Assessment for YOLO Object Detection: FID-Based Domain Shift Quantification and Hardware Profiling

Deploying a YOLO object detector trained on daytime images into a nighttime surveillance environment degrades mAP by 15–40%, yet most MLOps pipelines report only in-distribution accuracy. We present an automated domain shift assessment module that quantifies the Fréchet Inception Distance (FID) between training and deployment image distributions, profiles hardware complexity (GFLOPs, parameters, inference latency, peak VRAM), and exports publication-ready LaTeX comparison tables. The module executes as three steps inside a 15-step post-training forensic pipeline. Evaluated on a 250k-image industrial defect dataset across three domain pairs and four YOLO variants, FID ranges from 43.8 to 127.6, showing strong correlation with mAP degradation (Pearson r=0.962, p=5.9e-7).