LiteStyle_v1 data

LiteStyle: On-Device Fashion AnalyticsThe ProblemMobile e-commerce and digital wardrobes require real-time apparel analysis, but existing solutions often rely on GPU acceleration or cloud offloading, raising concerns regarding latency, privacy, and cost.Our SolutionLiteStyle is a lightweight framework for joint garment detection and formality prediction designed specifically for CPU-only devices.Architecture: Uses a YOLOv11n backbone. It reuses features from the SPPF module to power a compact StyleMLP head, eliminating the need for a secondary feature extractor.Deployment: Converts models to ONNX/OpenVINO with INT8 quantization. A Sequential Inference Protocol ensures memory stability across mobile SoCs and legacy x86 hardware.Key ResultsTested via generative-grounded synthetic imagery, LiteStyle delivers high-performance metrics on consumer-grade CPUs:Detection: 0.825 mAP@0.5 (DeepFashion2 subset).Classification: 0.921 AUC-ROC and 0.831 accuracy for formality.Efficiency: Achieves real-time throughput without GPU requirements.ContributionsLiteStyle provides a reproducible blueprint for CPU-centric fashion AI. Future work will expand the style taxonomy and move beyond synthetic benchmarks to real-world photograph datasets.