Implementation:Tencent Ncnn YOLOv3 Example

Knowledge Sources	Tencent_Ncnn
Domains	Vision, Object Detection
Last Updated	2026-02-09 19:00 GMT

Overview

Concrete tool for object detection on PASCAL VOC classes using YOLOv3 with a MobileNetV2-YOLO backbone and ncnn.

Description

This example demonstrates YOLOv3 object detection using a MobileNetV2 backbone (MobileNetV2-YOLO) with the ncnn framework. It loads the pretrained model, resizes input images to 352x352, and applies mean subtraction (127.5) followed by normalization (0.007843 = 1/127.5). The model extracts the "detection_out" layer and parses detection results in the standard (label, prob, x1, y1, x2, y2) format for the 20 PASCAL VOC classes. Unlike the YOLOv2 example, this uses a single-step normalization call rather than the two-step Caffe-YOLOv2 approach.

Usage

Use this example for YOLOv3-based object detection when a lightweight MobileNetV2 backbone is preferred. It offers improved accuracy over the YOLOv2 MobileNet variant while maintaining efficient inference on mobile hardware.

Code Reference

Source Location

• Repository: Tencent_Ncnn
• File: examples/yolov3.cpp
• Lines: 1-144

Signature

static int detect_yolov3(const cv::Mat& bgr, std::vector<Object>& objects);
static void draw_objects(const cv::Mat& bgr, const std::vector<Object>& objects);
int main(int argc, char** argv);

Import

#include "net.h"

I/O Contract

Inputs

Name	Type	Required	Description
image_path	const char*	Yes	Path to input image (passed as command-line argument)
bgr	const cv::Mat&	Yes	BGR image loaded by OpenCV, resized internally to 352x352

Outputs

Name	Type	Description
objects	std::vector<Object>	Detected objects with label, prob, and rect for VOC classes
Visual output	cv::imshow window	Displayed image with bounding boxes and class labels

Model Files

File	Description
mobilenetv2_yolov3.param	MobileNetV2-YOLOv3 network parameter file
mobilenetv2_yolov3.bin	MobileNetV2-YOLOv3 network weight file

Preprocessing

• Resize: Input resized to 352x352 using ncnn::Mat::from_pixels_resize with PIXEL_BGR
• Mean values: [127.5, 127.5, 127.5]
• Norm values: [0.007843, 0.007843, 0.007843] (i.e., 1/127.5)
• Effect: Normalizes pixel values to the range [-1, 1]

Usage Examples

Running the Example

./yolov3 image.jpg

Key Code Pattern

ncnn::Net yolov3;
yolov3.opt.use_vulkan_compute = true;

yolov3.load_param("mobilenetv2_yolov3.param");
yolov3.load_model("mobilenetv2_yolov3.bin");

ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR,
    bgr.cols, bgr.rows, 352, 352);

const float mean_vals[3] = {127.5f, 127.5f, 127.5f};
const float norm_vals[3] = {0.007843f, 0.007843f, 0.007843f};
in.substract_mean_normalize(mean_vals, norm_vals);

ncnn::Extractor ex = yolov3.create_extractor();
ex.input("data", in);

ncnn::Mat out;
ex.extract("detection_out", out);

for (int i = 0; i < out.h; i++)
{
    const float* values = out.row(i);
    // values[0]=label, values[1]=prob, values[2..5]=box coords
}

Related Pages

• Environment:Tencent_Ncnn_Build_Environment