Evaluating the Impact of INT8 Post-Training Quantization on Small-Target Detection for Edge-Compute Devices

Deploying advanced computer vision models on Unmanned Aerial Vehicles (UAVs) requires aggressive optimization to overcome the strict memory and power limitations of edge-computing hardware. This paper evaluates the impact of 8-bit Integer (INT8) Post-Training Quantization on the YOLO11n architecture, specifically focusing on the degradation of small-target detection capabilities. Using the high-altitude VisDrone dataset, a baseline 32-bit floating-point (FP32) model was trained and subsequently quantized to an INT8 LiteRT format. The ablation study reveals a 71.4\% reduction in model size (from 10.14 MB to 2.90 MB) and an increase in inference speed. However, this compression caused the overall mean Average Precision (mAP50-95) to degrade from 16.53\% to 8.05\%. Class-wise analysis proves that while large targets like cars remained moderately viable, the model experienced catastrophic forgetting on low-pixel-density targets, with pedestrian detection accuracy plummeting to 3.48\%. These findings highlight the critical limitations of standard INT8 quantization for edge-based aerial surveillance.