Implementation:Tencent Ncnn YOLOv4 Example

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

Overview

Concrete tool for object detection on COCO classes using YOLOv4 or YOLOv4-tiny with ncnn, supporting both image and video input with optional profiling.

Description

This example demonstrates YOLOv4 (or YOLOv4-tiny) object detection with ncnn, and is the only example in the repository that supports video input via OpenCV VideoCapture. The model variant is selected via a compile-time #define YOLOV4_TINY: YOLOv4-tiny uses 416x416 input, while full YOLOv4 uses 608x608. The model detects 80 COCO classes (person, car, dog, etc.). A separate init_yolov4 function handles model initialization, allowing the model to be loaded once and reused across video frames. The example also includes NCNN_PROFILING support via ncnn::get_current_time() for benchmarking model load time, detection time, capture time, and draw time. For video input, the detection loop runs continuously reading from a V4L2 device path.

Usage

Use this example for real-time object detection from a webcam or video device, or for single-image detection on COCO categories. It is particularly suitable for performance benchmarking and streaming inference scenarios. The YOLOv4-tiny variant is appropriate for real-time applications, while full YOLOv4 provides higher accuracy.

Code Reference

Source Location

• Repository: Tencent_Ncnn
• File: examples/yolov4.cpp
• Lines: 1-284

Signature

static int init_yolov4(ncnn::Net* yolov4, int* target_size);
static int detect_yolov4(const cv::Mat& bgr, std::vector<Object>& objects,
    int target_size, ncnn::Net* yolov4);
static int draw_objects(const cv::Mat& bgr, const std::vector<Object>& objects,
    int is_streaming);
int main(int argc, char** argv);

Import

#include "net.h"
#include "benchmark.h"  // for NCNN_PROFILING

I/O Contract

Inputs

Name	Type	Required	Description
device_or_image	const char*	Yes	Either a V4L2 device path (e.g., /dev/video0) for streaming or an image file path
bgr	cv::Mat	Yes	BGR frame from image file or video capture

Outputs

Name	Type	Description
objects	std::vector<Object>	Detected objects with label, prob, and rect for 80 COCO classes
Visual output	cv::imshow window	Image/video with bounding boxes and class labels (waitKey(1) for streaming, waitKey(0) for image)
Profiling output	stdout	Timing for model init, capture, detection, and draw (when NCNN_PROFILING is defined)

Model Files

File	Description
yolov4-tiny-opt.param	YOLOv4-tiny optimized parameter file (when YOLOV4_TINY defined)
yolov4-tiny-opt.bin	YOLOv4-tiny optimized weight file (when YOLOV4_TINY defined)
yolov4-opt.param	Full YOLOv4 optimized parameter file (when YOLOV4_TINY not defined)
yolov4-opt.bin	Full YOLOv4 optimized weight file (when YOLOV4_TINY not defined)

Preprocessing

• Color conversion: BGR to RGB via ncnn::Mat::PIXEL_BGR2RGB
• Resize: 416x416 for YOLOv4-tiny, 608x608 for full YOLOv4
• Mean values: [0, 0, 0]
• Norm values: [1/255.0, 1/255.0, 1/255.0]
• Effect: Normalizes pixel values to [0, 1] range

Architecture

The example separates initialization from detection to support the video streaming loop:

main()
  |-- init_yolov4()          // Load model once
  |-- [Image mode]
  |     |-- detect_yolov4()  // Single inference
  |     |-- draw_objects()   // Display and exit
  |-- [Video mode]
        |-- loop:
              |-- cap >> frame
              |-- detect_yolov4()
              |-- draw_objects(is_streaming=1)

Usage Examples

Running with an Image

./yolov4 image.jpg

Running with a Webcam

./yolov4 /dev/video0

Key Code Pattern

ncnn::Net yolov4;
int target_size = 0;
init_yolov4(&yolov4, &target_size);

// For video streaming
cv::VideoCapture cap;
cap.open(devicepath);

while (1) {
    cap >> frame;

    detect_yolov4(frame, objects, target_size, &yolov4);
    draw_objects(frame, objects, /*is_streaming=*/1);
}

Compile-Time Configuration

Define	Effect
YOLOV4_TINY	Selects YOLOv4-tiny model (416x416) instead of full YOLOv4 (608x608)
NCNN_PROFILING	Enables timing measurements for init, capture, detection, and draw phases

Related Pages

• Environment:Tencent_Ncnn_Build_Environment