Implementation:Tencent Ncnn SimplePose Example
| Knowledge Sources | |
|---|---|
| Domains | Vision, Pose Estimation |
| Last Updated | 2026-02-09 19:00 GMT |
Overview
Concrete tool for human pose estimation using the Simple Baseline model from GluonCV with ncnn, detecting 17 body keypoints via heatmap regression.
Description
This example demonstrates top-down human pose estimation using the Simple Baseline model (originally exported from MXNet via mxnet2ncnn). It loads the pose model, resizes input to 192x256 with ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), and extracts heatmaps from the "conv3_fwd" output layer. For each of the 17 body keypoints, it finds the maximum activation in the corresponding heatmap channel and maps it back to original image coordinates. The visualization draws both joint points (green circles) and skeleton connections (blue lines) for 16 predefined bone pairs. Keypoints with probability below 0.2 are filtered out.
Usage
Use this example for single-person pose estimation from a cropped image of a person. It is suitable for fitness tracking, gesture recognition, or action analysis applications. The model expects a pre-cropped person image; for multi-person scenarios, combine with a person detector first.
Code Reference
Source Location
- Repository: Tencent_Ncnn
- File: examples/simplepose.cpp
- Lines: 1-156
Signature
static int detect_posenet(const cv::Mat& bgr, std::vector<KeyPoint>& keypoints);
static void draw_pose(const cv::Mat& bgr, const std::vector<KeyPoint>& keypoints);
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 of a person (passed as command-line argument) |
| bgr | const cv::Mat& | Yes | BGR image loaded by OpenCV, converted to RGB and resized to 192x256 |
Outputs
| Name | Type | Description |
|---|---|---|
| keypoints | std::vector<KeyPoint> | 17 detected body keypoints, each with a cv::Point2f position and float probability |
| Visual output | cv::imshow window | Image with green joint circles and blue skeleton lines (threshold > 0.2) |
Model Files
| File | Description |
|---|---|
| pose.param | Simple Baseline pose model parameter file |
| pose.bin | Simple Baseline pose model weight file |
Preprocessing
- Color conversion: BGR to RGB via
ncnn::Mat::PIXEL_BGR2RGB - Resize: Input resized to 192x256
- Mean values: [0.485*255, 0.456*255, 0.406*255] = [123.675, 116.28, 103.53]
- Norm values: [1/0.229/255, 1/0.224/255, 1/0.225/255] = [0.01712, 0.01751, 0.01742]
Keypoint Layout
The model detects 17 COCO-format keypoints with the following skeleton connections (16 bone pairs):
static const int joint_pairs[16][2] = {
{0, 1}, {1, 3}, {0, 2}, {2, 4}, {5, 6}, {5, 7},
{7, 9}, {6, 8}, {8, 10}, {5, 11}, {6, 12}, {11, 12},
{11, 13}, {12, 14}, {13, 15}, {14, 16}
};
Usage Examples
Running the Example
./simplepose person.jpg
Key Code Pattern
ncnn::Net posenet;
posenet.opt.use_vulkan_compute = true;
posenet.load_param("pose.param");
posenet.load_model("pose.bin");
ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB,
w, h, 192, 256);
const float mean_vals[3] = {0.485f * 255.f, 0.456f * 255.f, 0.406f * 255.f};
const float norm_vals[3] = {1 / 0.229f / 255.f, 1 / 0.224f / 255.f, 1 / 0.225f / 255.f};
in.substract_mean_normalize(mean_vals, norm_vals);
ncnn::Extractor ex = posenet.create_extractor();
ex.input("data", in);
ncnn::Mat out;
ex.extract("conv3_fwd", out);
// Resolve keypoints from heatmap channels
for (int p = 0; p < out.c; p++)
{
const ncnn::Mat m = out.channel(p);
// Find max activation location in heatmap
}