Weird output of resize_bilinear

[kyuusaku]

The function resize_bilinear_image in Mat.cpp seems not accurate.

The unit test reveals unexpected negative outputs for positive inputs.

Here is a simple example can be used to reproduce the weird output.

#include <iostream>
#include <fstream>
#include <opencv2/opencv.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include "net.h"

void ncnn_blob_to_txt(const ncnn::Mat& m, const std::string blob_name)
{
    std::ofstream outfile("log/" + blob_name + ".txt");
    if (!outfile.is_open())
    {
        std::cout << "file open fail" << std::endl;
        exit(1);
    }
    
    std::cout << "dims:" << m.dims << " w:" << m.w << " h:" << m.h << " c:" << m.c << std::endl;
    outfile << "dims:" << m.dims << " w:" << m.w << " h:" << m.h << " c:" << m.c << "\r\n" << " \r\n";

    int size = m.w * m.h;
    for (int c = 0; c < m.c; c++)
    {
        const float* ptr = m.channel(c);
        for (int h = 0; h < m.h; h++)
        {
            for (int w = 0; w < m.w; w++)
            {
                outfile << *ptr << " ";
                ptr++;
            }
            outfile << "\r\n";
        }
        outfile << " \r\n";
    }
    outfile.close();
}

void cv_blob_to_txt(const cv::Mat& m, const std::string blob_name)
{
    std::ofstream outfile("log/" + blob_name + ".txt");
    if (!outfile.is_open())
    {
        std::cout << "file open fail" << std::endl;
        exit(1);
    }
    
    std::cout << "dims:" << m.dims << " w:" << m.cols << " h:" << m.rows << std::endl;
    outfile << "dims:" << m.dims << " w:" << m.cols << " h:" << m.rows << "\r\n" << " \r\n";

    const float* ptr = (const float*)m.data;
    for (int h = 0; h < m.rows; h++)
    {
        for (int w = 0; w < m.cols; w++)
        {
            outfile << *(ptr + h*m.cols + w) << " ";
        }
        outfile << "\r\n";
    }
    outfile.close();
}

void ncnn_from_float(ncnn::Mat& m, float* input)
{
    float* ptr = m;

    int size = m.w * m.h;

    printf("%d %d %d \n", m.w, m.h, m.c);

    for (int h = 0; h < m.h; h++)
    {
        for (int w = 0; w < m.w; w++)
        {
            *(ptr + h*m.w + w) = *(input + h*m.w + w);
        }
    }
}

void cv_from_float(cv::Mat& m, float* input)
{
    float* ptr = (float*)m.data;

    printf("%d %d \n", m.cols, m.rows);

    for (int h = 0; h < m.rows; h++)
    {
        for (int w = 0; w < m.cols; w++)
        {
            *(ptr + h*m.cols + w) = *(input + h*m.cols + w);
        }
    }
}

int test_resize_bilinear(int argc, char** argv) {
    float input[12][12] = {
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4.81992e-14, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3.63482e-15, 4.02549e-13}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, 
        {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
    };


    ncnn::Mat in(12, 12, 1);
    ncnn_from_float(in, &input[0][0]);

    ncnn::Mat out;
    ncnn::resize_bilinear(in, out, 48, 48);

    ncnn_blob_to_txt(in, "test_resize_bilinear_input");
    ncnn_blob_to_txt(out, "test_resize_bilinear_output");

    cv::Mat cv_in(12, 12, CV_32FC1);
    cv_from_float(cv_in, &input[0][0]);
    cv::Mat cv_out;
    cv::resize(cv_in, cv_out, cv::Size(48, 48));
    cv_blob_to_txt(cv_out, "test_resize_bilinear_opencv_output");

    return 0;
}


int main(int argc, char* argv[]) {
    test_resize_bilinear(argc, argv);
}

The weird output is

dims:3 w:48 h:48 c:1
 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 -2.25934e-15 -6.77801e-15 -1.12967e-14 -1.58154e-14 -1.58154e-14 -1.12967e-14 -6.77801e-15 -2.25934e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 -7.53113e-16 -2.25934e-15 -3.76556e-15 -5.27179e-15 -5.27179e-15 -3.76556e-15 -2.25934e-15 -7.53113e-16 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 7.53113e-16 2.25934e-15 3.76556e-15 5.27179e-15 5.27179e-15 3.76556e-15 2.25934e-15 7.53113e-16 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2.25934e-15 6.77801e-15 1.12967e-14 1.58154e-14 1.58154e-14 1.12967e-14 6.77801e-15 2.25934e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3.76556e-15 1.12967e-14 1.88278e-14 2.63589e-14 2.63589e-14 1.88278e-14 1.12967e-14 3.76556e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5.27179e-15 1.58154e-14 2.63589e-14 3.69025e-14 3.69025e-14 2.63589e-14 1.58154e-14 5.27179e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5.27179e-15 1.58154e-14 2.63589e-14 3.69025e-14 3.69025e-14 2.63589e-14 1.58154e-14 5.27179e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3.76556e-15 1.12967e-14 1.88278e-14 2.63589e-14 2.63589e-14 1.88278e-14 1.12967e-14 3.76556e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2.25934e-15 6.77801e-15 1.12967e-14 1.58154e-14 1.58154e-14 1.12967e-14 6.77801e-15 2.25934e-15 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 7.53113e-16 2.25934e-15 3.76556e-15 5.27179e-15 5.27179e-15 3.76556e-15 2.25934e-15 7.53113e-16 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5.67941e-17 1.70382e-16 2.8397e-16 3.97558e-16 6.68739e-15 1.91535e-14 3.16195e-14 4.40856e-14 5.03186e-14 5.03186e-14 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1.70382e-16 5.11147e-16 8.51911e-16 1.19268e-15 2.00622e-14 5.74604e-14 9.48586e-14 1.32257e-13 1.50956e-13 1.50956e-13 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2.8397e-16 8.51911e-16 1.41985e-15 1.98779e-15 3.34369e-14 9.57673e-14 1.58098e-13 2.20428e-13 2.51593e-13 2.51593e-13 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3.97558e-16 1.19268e-15 1.98779e-15 2.78291e-15 4.68117e-14 1.34074e-13 2.21337e-13 3.08599e-13 3.5223e-13 3.5223e-13 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3.97558e-16 1.19268e-15 1.98779e-15 2.78291e-15 4.68117e-14 1.34074e-13 2.21337e-13 3.08599e-13 3.5223e-13 3.5223e-13 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2.8397e-16 8.51911e-16 1.41985e-15 1.98779e-15 3.34369e-14 9.57673e-14 1.58098e-13 2.20428e-13 2.51593e-13 2.51593e-13 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1.70382e-16 5.11147e-16 8.51911e-16 1.19268e-15 2.00622e-14 5.74604e-14 9.48586e-14 1.32257e-13 1.50956e-13 1.50956e-13 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5.67941e-17 1.70382e-16 2.8397e-16 3.97558e-16 6.68739e-15 1.91535e-14 3.16195e-14 4.40856e-14 5.03186e-14 5.03186e-14 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 

It is worth noting that. There are also other cases can generate such weird outputs.

@nihui please help.

[nihui]

confirmed