有没有办法在 OpenCV 中读取 .pfm 文件？
PFM 是一种不常见的图像格式，我不知道为什么 Middlebury 数据集选择使用它，可能是因为它使用浮点值。 无论如何，我能够使用 OpenCV 读取图像：
import numpy as np
import cv2
groundtruth = cv2.imread('disp0.pfm', cv2.IMREAD_UNCHANGED)
注意
IMREAD_UNCHANGED但是等一下：
inf# Remove infinite value to display
groundtruth[groundtruth==np.inf] = 0
# Normalize and convert to uint8
groundtruth = cv2.normalize(groundtruth, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8)
# Show
cv2.imshow("groundtruth", groundtruth)
cv2.waitKey(0)
cv2.destroyAllWindows()
根据“.pfm”文件格式的描述（参见http://netpbm.sourceforge.net/doc/pfm.html），我写了以下只依赖于标准C/C++的读写函数图书馆。它被证明可以很好地读取/写入 pfm 文件，例如 MiddleBury Computer Vision 的地面实况差异“.pfm”文件（参见http://vision.middlebury.edu/stereo/submit3/）。
#ifndef _PGM_H_
#define _PGM_H_
#include <fstream>
#include <iostream>
#include <algorithm>
#include <string>
#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <bitset> /*std::bitset<32>*/
#include <cstdio>
enum PFM_endianness { BIG, LITTLE, ERROR};
class PFM {
public:
PFM();
inline bool is_little_big_endianness_swap(){
if (this->endianess == 0.f) {
std::cerr << "this-> endianness is not assigned yet!\n";
exit(0);
}
else {
uint32_t endianness = 0xdeadbeef;
//std::cout << "\n" << std::bitset<32>(endianness) << std::endl;
unsigned char * temp = (unsigned char *)&endianness;
//std::cout << std::bitset<8>(*temp) << std::endl;
PFM_endianness endianType_ = ((*temp) ^ 0xef == 0 ?
LITTLE : (*temp) ^ (0xde) == 0 ? BIG : ERROR);
// ".pfm" format file specifies that:
// positive scale means big endianess;
// negative scale means little endianess.
return ((BIG == endianType_) && (this->endianess < 0.f))
|| ((LITTLE == endianType_) && (this->endianess > 0.f));
}
}
template<typename T>
T * read_pfm(const std::string & filename) {
FILE * pFile;
pFile = fopen(filename.c_str(), "rb");
char c[100];
if (pFile != NULL) {
fscanf(pFile, "%s", c);
// strcmp() returns 0 if they are equal.
if (!strcmp(c, "Pf")) {
fscanf(pFile, "%s", c);
// atoi: ASCII to integer.
// itoa: integer to ASCII.
this->width = atoi(c);
fscanf(pFile, "%s", c);
this->height = atoi(c);
int length_ = this->width * this->height;
fscanf(pFile, "%s", c);
this->endianess = atof(c);
fseek(pFile, 0, SEEK_END);
long lSize = ftell(pFile);
long pos = lSize - this->width*this->height * sizeof(T);
fseek(pFile, pos, SEEK_SET);
T* img = new T[length_];
//cout << "sizeof(T) = " << sizeof(T);
fread(img, sizeof(T), length_, pFile);
fclose(pFile);
/* The raster is a sequence of pixels, packed one after another,
* with no delimiters of any kind. They are grouped by row,
* with the pixels in each row ordered left to right and
* the rows ordered bottom to top.
*/
T* tbimg = (T *)malloc(length_ * sizeof(T));// top-to-bottom.
//PFM SPEC image stored bottom -> top reversing image
for (int i = 0; i < this->height; i++) {
memcpy(&tbimg[(this->height - i - 1)*(this->width)],
&img[(i*(this->width))],
(this->width) * sizeof(T));
}
if (this->is_little_big_endianness_swap()){
std::cout << "little-big endianness transformation is needed.\n";
// little-big endianness transformation is needed.
union {
T f;
unsigned char u8[sizeof(T)];
} source, dest;
for (int i = 0; i < length_; ++i) {
source.f = tbimg[i];
for (unsigned int k = 0, s_T = sizeof(T); k < s_T; k++)
dest.u8[k] = source.u8[s_T - k - 1];
tbimg[i] = dest.f;
//cout << dest.f << ", ";
}
}
delete[] img;
return tbimg;
}
else {
std::cout << "Invalid magic number!"
<< " No Pf (meaning grayscale pfm) is missing!!\n";
fclose(pFile);
exit(0);
}
}
else {
std::cout << "Cannot open file " << filename
<< ", or it does not exist!\n";
fclose(pFile);
exit(0);
}
}
template<typename T>
void write_pfm(const std::string & filename, const T* imgbuffer,
const float & endianess_) {
std::ofstream ofs(filename.c_str(), std::ifstream::binary);
// ** 1) Identifier Line: The identifier line contains the characters
// "PF" or "Pf". PF means it's a color PFM.
// Pf means it's a grayscale PFM.
// ** 2) Dimensions Line:
// The dimensions line contains two positive decimal integers,
// separated by a blank. The first is the width of the image;
// the second is the height. Both are in pixels.
// ** 3) Scale Factor / Endianness:
// The Scale Factor / Endianness line is a queer line that jams
// endianness information into an otherwise sane description
// of a scale. The line consists of a nonzero decimal number,
// not necessarily an integer. If the number is negative, that
// means the PFM raster is little endian. Otherwise, it is big
// endian. The absolute value of the number is the scale
// factor for the image.
// The scale factor tells the units of the samples in the raster.
// You use somehow it along with some separately understood unit
// information to turn a sample value into something meaningful,
// such as watts per square meter.
ofs << "Pf\n"
<< this->width << " " << this->height << "\n"
<< endianess_ << "\n";
/* PFM raster:
* The raster is a sequence of pixels, packed one after another,
* with no delimiters of any kind. They are grouped by row,
* with the pixels in each row ordered left to right and
* the rows ordered bottom to top.
* Each pixel consists of 1 or 3 samples, packed one after another,
* with no delimiters of any kind. 1 sample for a grayscale PFM
* and 3 for a color PFM (see the Identifier Line of the PFM header).
* Each sample consists of 4 consecutive bytes. The bytes represent
* a 32 bit string, in either big endian or little endian format,
* as determined by the Scale Factor / Endianness line of the PFM
* header. That string is an IEEE 32 bit floating point number code.
* Since that's the same format that most CPUs and compiler use,
* you can usually just make a program use the bytes directly
* as a floating point number, after taking care of the
* endianness variation.
*/
int length_ = this->width*this->height;
this->endianess = endianess_;
T* tbimg = (T *)malloc(length_ * sizeof(T));
// PFM SPEC image stored bottom -> top reversing image
for (int i = 0; i < this->height; i++) {
memcpy(&tbimg[(this->height - i - 1)*this->width],
&imgbuffer[(i*this->width)],
this->width * sizeof(T));
}
if (this->is_little_big_endianness_swap()) {
std::cout << "little-big endianness transformation is needed.\n";
// little-big endianness transformation is needed.
union {
T f;
unsigned char u8[sizeof(T)];
} source, dest;
for (int i = 0; i < length_; ++i) {
source.f = tbimg[i];
for (size_t k = 0, s_T = sizeof(T); k < s_T; k++)
dest.u8[k] = source.u8[s_T - k - 1];
tbimg[i] = dest.f;
//cout << dest.f << ", ";
}
}
ofs.write((char *)tbimg, this->width*this->height * sizeof(T));
ofs.close();
free(tbimg);
}
inline float getEndianess(){return endianess;}
inline int getHeight(void){return height;}
inline int getWidth(void){return width;}
inline void setHeight(const int & h){height = h;}
inline void setWidth(const int & w){width = w;}
private:
int height;
int width;
float endianess;
};
#endif /* PGM_H_ */
请原谅我在代码中留下了很多无用的注释。
一个简单的例子显示写/读：
int main(){
PFM pfm_rw;
string temp = "img/Motorcycle/disp0GT.pfm";
float * p_disp_gt = pfm_rw.read_pfm<float>(temp);
//int imgH = pfm_rw.getHeight();
//int imgW = pfm_rw.getWidth();
//float scale = pfm_rw.getEndianess();
string temp2 = "result/Motorcycle/disp0GT_n1.pfm";
pfm_rw.write_pfm<float>(temp2, p_disp_gt, -1.0f);
return 1;
}
据我所知，OpenCV不支持直接读取PFM文件。
您可以参考代码片段这里一个简单的PFM阅读器，这将使您能够将PFM文件读入
COLOR *dataCOLORtypedef struct {
float r;
float g;
float b;
} COLOR;