Cross-Regional Benchmarking of Enterprise Cloud Object Recognition Services: A Comparative Analysis of AWS, Azure, Google Cloud Vision, and YOLO12

This paper presents a benchmark comparing four object recognition systems, namely AWS Rekognition, Azure Computer Vision, Google Cloud Vision, and YOLO12, using the GeoDE dataset, a geographically diverse benchmark containing 61,940 crowd-sourced images across 40 object classes and 6 world regions. We evaluate performance using contains-label accuracy, precision, F1-score, and mean confidence scores on a pre-specified subset of 12 of the 40 categories (30%), selected because their labels occur broadly across the included countries and regions and support cross-system comparison. On this constrained subset, AWS Rekognition achieves the highest accuracy (94.49%, 95% CI: [93.23%, 95.51%]) with significantly superior performance compared to Azure (p<0.001, Cohen’s d=0.65) and Google (p<0.001, d=0.72). This aggregate accuracy is specific to the selected categories and should not be interpreted as performance on all GeoDE categories or general object recognition. Precision analysis indicates this lead is not explained by a noisier positive label set: AWS achieved 1.000 weighted precision, compared with 0.9997 for Azure and 0.9998 for Google. When restricted to vocabulary-overlapping categories, YOLO12 demonstrates competitive accuracy (95.67%) within 4% of cloud services. Its measured local CPU inference time was 12.3 ms, whereas the cloud measurements were end-to-end API round-trip times, including network overhead, of approximately 780 ms (Azure), 850 ms (AWS), and 920 ms (Google) in our setup; therefore, these figures are not a like-for-like comparison of model processing latency. While all systems show descriptively higher accuracy on European content, Kruskal–Wallis tests indicate regional differences are not statistically significant at α=0.05, though effect sizes (d=0.13–0.61) suggest meaningful practical disparities warranting further investigation. As expected from the pre-specified category design, culturally variable objects (stall, lighter) show substantially lower recognition rates than standardized objects (car, truck). Human baseline evaluation on a stratified subset (n=500) yielded 97.2% accuracy with inter-annotator agreement (Fleiss’ κ=0.94), indicating algorithmic performance approaches but does not reach human-level performance for most categories. These findings are useful to practitioners selecting vision services, cloud providers improving geographically robust products, and researchers studying fairness in computer vision.