Skip to content
Conference

A study on lightweight YOLOv10 based on mixed-precision post-training dynamic quantization for edge deployment

Nov 2026 · International Conference on Algorithms, Image Processing, and Deep Learning (AIPDL 2026) · 0 citations

Abstract

As an advanced real-time end-to-end object detection model, YOLOv10 faces challenges of large model size and high computational overhead when deployed on edge devices. To address this issue, this paper proposes a lightweight method based on post-training mixed-precision quantization using INT8. Using YOLOv10n as the baseline model, training is first performed on the COCO128 and PASCAL VOC datasets; Subsequently, by exporting the model in ONNX format, we utilize the `quantize_dynamic` interface of ONNX Runtime to perform dynamic quantization. This involves storing convolutional layer weights in INT8 format and performing runtime dynamic quantization on activation values, while retaining FP32 precision for the detection heads (Classification Head and Regression Head) to maintain localization and classification accuracy. Experimental results show that on the PASCAL VOC dataset, the model size was reduced from 9.0MB to 2.5MB after quantization, achieving a compression ratio of 3.6x;CPU inference speed increased from 35FPS to 85FPS, with a 2.4x speedup; meanwhile, the mean average precision (mAP50-95) dropped only from 53.1% to 51.2%, with accuracy loss controlled within 1.9 percentage points. This method requires no retraining or data calibration and can be directly applied to pre-trained models, providing a practical deployment solution for the efficient deployment of YOLOv10 in resource-constrained scenarios.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.