Cross-Domain Generalization Assessment for Object Detection Models: FID-Based Domain Shift Detection and Edge Hardware Profiling

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Deploying deep learning object detectors into unconstrained operational environments inevitably exposes vision models to distributional domain shift. Variations in ambient illumination, sensor noise profiles, meteorological conditions, and geometric viewpoints trigger severe performance degradation that is difficult to forecast prior to deployment. In this paper, we propose an automated post-training framework that integrates Fréchet Inception Distance (FID) for distribution divergence auditing with comprehensive hardware complexity profiling for edge accelerators. Across 12 cross-domain shift combinations spanning synthetic rendering, daytime illumination, nighttime conditions, and adverse rain, our empirical findings reveal a strong linear correlation (r=0.74) between latent feature FID and empirical mean Average Precision (mAP) drop. Specifically, domain shifts with FID > 120 consistently result in catastrophic accuracy degradation exceeding 35%. Furthermore, our automated hardware profiler maps the pareto frontier across model scales (YOLO-n, YOLO-s, YOLO-m) and spatial resolutions (320px to 1280px), demonstrating that quadrupling input resolution increases computational demand fourfold while introducing a 3.3x latency penalty. Enforcing an automated gating threshold at FID <= 100 successfully filters out 38% of hazardous deployments, lowering edge production failure rates from 42% to below 5%.