Implementation:Tencent Ncnn DFL Anchor Free Decode

Knowledge Sources	ncnn
Domains	Computer_Vision, Object_Detection
Last Updated	2026-02-09 00:00 GMT

Overview

Pattern implementation for decoding DFL-based anchor-free detection outputs into bounding boxes, as demonstrated in ncnn YOLOv8 examples.

Description

This is a Pattern Doc documenting the DFL bounding box decoding pattern from the ncnn yolov8.cpp example. The implementation uses ncnn::create_layer("Softmax") internally to compute DFL softmax over the 16-bin distribution for each box edge. The pattern processes the raw detection output tensor (shape: w=num_classes+64, h=num_anchors) and produces a vector of decoded Object structs.

The implementation handles three detection strides (8, 16, 32) and uses a confidence threshold to filter low-probability detections before decoding, saving computation.

Usage

Implement this pattern for YOLOv8 and YOLO11 anchor-free detection post-processing. Adapt reg_max and stride values for different model architectures.

Code Reference

Source Location

• Repository: ncnn
• File: examples/yolov8.cpp
• Lines: L160-254 (generate_proposals with DFL decode), L256-273 (multi-stride dispatch)

Signature

// Pattern: DFL bounding box decode
// User-implemented function based on ncnn example pattern

struct Object
{
    cv::Rect_<float> rect;  // bounding box (x, y, width, height)
    int label;               // class index
    float prob;              // confidence score
};

static void generate_proposals(
    const ncnn::Mat& pred,   // raw detection output
    int stride,              // detection stride (8, 16, or 32)
    const ncnn::Mat& in_pad, // letterboxed input (for shape)
    float prob_threshold,    // confidence threshold (e.g., 0.25)
    std::vector<Object>& objects  // output: decoded detections
);

Import

#include "layer.h"  // for ncnn::create_layer("Softmax")
#include "net.h"
#include <vector>

I/O Contract

Inputs

Name	Type	Required	Description
pred	ncnn::Mat	Yes	Raw detection output (w=num_class+reg_max*4, h=num_anchors)
stride	int	Yes	Detection stride (8, 16, or 32)
prob_threshold	float	Yes	Minimum confidence threshold (e.g., 0.25)

Outputs

Name	Type	Description
objects	std::vector<Object>	Decoded detections with rect, label, and prob fields

Usage Examples

YOLOv8 Multi-Stride Decode

// After inference: extract single output blob
ncnn::Mat out;
ex.extract("out0", out);

// Decode at three strides
std::vector<Object> proposals;
std::vector<Object> objects8, objects16, objects32;

// Stride 8 proposals
{
    ncnn::Mat pred(out.w, out.h / 3, (float*)out.data);
    generate_proposals(pred, 8, in_pad, 0.25f, objects8);
    proposals.insert(proposals.end(), objects8.begin(), objects8.end());
}
// Stride 16 and 32 similarly...

// Apply NMS to merged proposals

Related Pages

Implements Principle

• Principle:Tencent_Ncnn_Bounding_Box_Decoding