Real-Time Quantized YOLO Object Detection on Serverless Cloud Functions:An Experimental and Analytical Study

Serverless computing has emerged as an attractive execution model for event-driven applicationsdue to its automatic scalability, fine-grained billing, and minimal operational overhead . Theseproperties make serverless platforms appealing for machine learning inference workloads with variableor bursty demand. However, deploying real-time computer vision pipelines on serverless infrastructuresremains challenging due to CPU-only execution, memory constraints, and cold-start overheads [5, 6].This paper presents a systematic experimental and analytical study of quantized object detectiondeployed on serverless cloud functions. Using quantized YOLO-based models executed on AWSLambda with ONNX Runtime , we investigate latency–cost trade-offs under practical deploy-ment constraints. Experimental results show that INT8 quantization substantially reduces warm-startinference latency , while cold-start behavior remains the dominant bottleneck.