Implementation:Tencent Ncnn PeleeNetSSD Seg Example

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

Overview

Concrete tool for combined object detection and semantic segmentation using PeleeNet-SSD with ncnn in a single forward pass.

Description

This example demonstrates a multi-task model that performs both object detection and semantic segmentation simultaneously. It loads a PeleeNet model (converted from MobileNet-YOLO), resizes input to 304x304, and extracts two outputs: the "detection_out" layer for bounding box detections and the "sigmoid" layer for a segmentation map. The segmentation map is resized back to original image dimensions using bilinear interpolation and overlaid as a colored mask. Detection classes include driving-related categories: person, rider, car, bus, truck, bike, motor, traffic light, traffic sign, and train. Segmentation pixels exceeding a 0.45 threshold are blended with the original image using color mapping.

Usage

Use this example for autonomous driving or traffic scene understanding where both object detection and road segmentation are needed from a single efficient model. It is suitable for scenarios requiring simultaneous perception of discrete objects and continuous regions.

Code Reference

Source Location

• Repository: Tencent_Ncnn
• File: examples/peleenetssd_seg.cpp
• Lines: 1-187

Signature

static int detect_peleenet(const cv::Mat& bgr, std::vector<Object>& objects, ncnn::Mat& resized);
static void draw_objects(const cv::Mat& bgr, const std::vector<Object>& objects, ncnn::Mat map);
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 304x304

Outputs

Name	Type	Description
objects	std::vector<Object>	Detected objects with label, prob, and rect for driving-related classes
resized	ncnn::Mat	Segmentation map resized to original image dimensions via bilinear interpolation
Visual output	cv::imshow window	Image with bounding boxes, class labels, and colored segmentation overlay

Model Files

File	Description
pelee.param	PeleeNet network parameter file
pelee.bin	PeleeNet network weight file

Preprocessing

• Resize: Input resized to 304x304 using ncnn::Mat::from_pixels_resize with PIXEL_BGR
• Mean values: [103.9, 116.7, 123.6]
• Norm values: [0.017, 0.017, 0.017]
• Segmentation post-processing: Output from "sigmoid" layer resized via resize_bilinear to original image dimensions

Usage Examples

Running the Example

./peleenetssd_seg image.jpg

Key Code Pattern

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

peleenet.load_param("pelee.param");
peleenet.load_model("pelee.bin");

ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR,
    bgr.cols, bgr.rows, 304, 304);

const float mean_vals[3] = {103.9f, 116.7f, 123.6f};
const float norm_vals[3] = {0.017f, 0.017f, 0.017f};
in.substract_mean_normalize(mean_vals, norm_vals);

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

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

ncnn::Mat seg_out;
ex.extract("sigmoid", seg_out);    // Segmentation map
resize_bilinear(seg_out, resized, img_w, img_h);

Related Pages

• Environment:Tencent_Ncnn_Build_Environment