This paper focuses on bridging the gap between benchmark performance and real-world feasibility of object detectors on consumer-grade hardware. While models like YOLOv10s achieve real-time speeds, these performance metrics are typically achieved on high-performance desktop-grade GPUs. On resource-constrained systems like the RTX 4060 GPU, we demonstrate that system-level bottlenecks, rather than computational speed, are the primary cause of performance degradation. To address this, we present a two-pass adaptive inference algorithm that can be applied without changing the model architecture. This algorithm accelerates by leveraging a fast low-resolution pass and, when necessary, a high-resolution pass. We achieve a 1.85x speedup and a 5.51% mAP loss compared to the PyTorch early-exit baseline on the 5,000-image COCO dataset. Rather than relying on pure model optimization, we present a practical and reproducible approach to maximizing throughput through a hardware-aware inference strategy.
