apoorva/greenhouse
2
1
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Session: Compression of the vocabulary to avoid QByteArray max 2GB limitation

Browse Source
This commit is contained in:
matlabbe 2018-09-10 18:57:51 -04:00
parent c6cf667645
commit 556bf5f4cc
9 changed files with 259 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@ -64,6 +64,8 @@ ADD_DEFINITIONS(-DQT_NO_KEYWORDS) # To avoid conflicts with boost signals used i
FIND_PACKAGE(Tcmalloc QUIET)
FIND_PACKAGE(ZLIB REQUIRED QUIET)
SET(NONFREE 0)
IF(OPENCV_NONFREE_FOUND OR OPENCV_XFEATURES2D_FOUND)
SET(NONFREE 1)

3
src/CMakeLists.txt
View File

@ -74,6 +74,7 @@ SET(SRC_FILES
./utilite/UConversion.cpp
./rtabmap/PdfPlot.cpp
./json/jsoncpp.cpp
./Compression.cpp
${moc_srcs}
${moc_uis}
${srcs_qrc}
@ -93,6 +94,7 @@ SET(INCLUDE_DIRS
${CMAKE_CURRENT_SOURCE_DIR}
${OpenCV_INCLUDE_DIRS}
${CMAKE_CURRENT_BINARY_DIR} # for qt ui generated in binary dir
${ZLIB_INCLUDE_DIRS}
)
IF(CATKIN_BUILD)
SET(INCLUDE_DIRS
@ -108,6 +110,7 @@ ENDIF(QT4_FOUND)
SET(LIBRARIES
${QT_LIBRARIES}
${OpenCV_LIBS}
${ZLIB_LIBRARIES}
)
IF(CATKIN_BUILD)
SET(LIBRARIES

79
src/Compression.cpp Normal file
View File

@ -0,0 +1,79 @@
/*
* Compression.cpp
*
* Created on: Sep 10, 2018
* Author: labm2414
*/
#include <Compression.h>
#include <zlib.h>
#include "find_object/utilite/ULogger.h"
namespace find_object {
std::vector<unsigned char> compressData(const cv::Mat & data)
{
std::vector<unsigned char> bytes;
if(!data.empty())
{
uLong sourceLen = uLong(data.total())*uLong(data.elemSize());
uLong destLen = compressBound(sourceLen);
bytes.resize(destLen);
int errCode = compress(
(Bytef *)bytes.data(),
&destLen,
(const Bytef *)data.data,
sourceLen);
bytes.resize(destLen+3*sizeof(int));
*((int*)&bytes[destLen]) = data.rows;
*((int*)&bytes[destLen+sizeof(int)]) = data.cols;
*((int*)&bytes[destLen+2*sizeof(int)]) = data.type();
if(errCode == Z_MEM_ERROR)
{
UERROR("Z_MEM_ERROR : Insufficient memory.");
}
else if(errCode == Z_BUF_ERROR)
{
UERROR("Z_BUF_ERROR : The buffer dest was not large enough to hold the uncompressed data.");
}
}
return bytes;
}
cv::Mat uncompressData(const unsigned char * bytes, unsigned long size)
{
cv::Mat data;
if(bytes && size>=3*sizeof(int))
{
//last 3 int elements are matrix size and type
int height = *((int*)&bytes[size-3*sizeof(int)]);
int width = *((int*)&bytes[size-2*sizeof(int)]);
int type = *((int*)&bytes[size-1*sizeof(int)]);
data = cv::Mat(height, width, type);
uLongf totalUncompressed = uLongf(data.total())*uLongf(data.elemSize());
int errCode = uncompress(
(Bytef*)data.data,
&totalUncompressed,
(const Bytef*)bytes,
uLong(size));
if(errCode == Z_MEM_ERROR)
{
UERROR("Z_MEM_ERROR : Insufficient memory.");
}
else if(errCode == Z_BUF_ERROR)
{
UERROR("Z_BUF_ERROR : The buffer dest was not large enough to hold the uncompressed data.");
}
else if(errCode == Z_DATA_ERROR)
{
UERROR("Z_DATA_ERROR : The compressed data (referenced by source) was corrupted.");
}
}
return data;
}
} /* namespace find_object */

20
src/Compression.h Normal file
View File

@ -0,0 +1,20 @@
/*
* Compression.h
*
* Created on: Sep 10, 2018
* Author: labm2414
*/
#ifndef SRC_COMPRESSION_H_
#define SRC_COMPRESSION_H_
#include <opencv2/opencv.hpp>
namespace find_object {
std::vector<unsigned char> compressData(const cv::Mat & data);
cv::Mat uncompressData(const unsigned char * bytes, unsigned long size);
}
#endif /* SRC_COMPRESSION_H_ */

11
src/FindObject.cpp
View File

@ -735,7 +735,7 @@ protected:
{
QTime time;
time.start();
UINFO("Extracting descriptors from object %d...", objectId_);
UDEBUG("Extracting descriptors from object %d...", objectId_);
QTime timeStep;
timeStep.start();
@ -850,6 +850,7 @@ private:
void FindObject::updateObjects(const QList<int> & ids)
{
UINFO("Update %d objects...", ids.size());
QList<ObjSignature*> objectsList;
if(ids.size())
{
@ -884,7 +885,7 @@ void FindObject::updateObjects(const QList<int> & ids)
if(objectsList.size())
{
UINFO("Features extraction from %d objects...", objectsList.size());
UINFO("Features extraction from %d objects... (threads=%d)", objectsList.size(), threadCounts);
for(int i=0; i<objectsList.size(); i+=threadCounts)
{
QVector<ExtractFeaturesThread*> threads;
@ -995,6 +996,8 @@ void FindObject::updateVocabulary(const QList<int> & ids)
}
}
UINFO("Updating vocabulary with %d objects and %d descriptors...", ids.size(), count);
// Copy data
if(count)
{
@ -1100,6 +1103,10 @@ void FindObject::updateVocabulary(const QList<int> & ids)
}
if(addedWords && !Settings::getGeneral_vocabularyFixed())
{
if(!incremental)
{
UINFO("Updating vocabulary...");
}
vocabulary_->update();
}

21
src/MainWindow.cpp
View File

@ -244,15 +244,23 @@ MainWindow::MainWindow(FindObject * findObject, Camera * camera, QWidget * paren
if(findObject_->objects().size())
{
UINFO("Creating %d object widgets...", findObject_->objects().size());
// show objects already loaded in FindObject
int i=0;
for(QMap<int, ObjSignature *>::const_iterator iter = findObject_->objects().constBegin();
iter!=findObject_->objects().constEnd();
++iter)
{
ObjWidget * obj = new ObjWidget(iter.key(), iter.value()->keypoints(), iter.value()->words(), cvtCvMat2QImage(iter.value()->image()));
ObjWidget * obj = new ObjWidget(iter.key(), iter.value()->keypoints(), iter.value()->words(), iter.value()->image().empty()?QImage():cvtCvMat2QImage(iter.value()->image()));
objWidgets_.insert(obj->id(), obj);
this->showObject(obj);
++i;
if(i % 100 == 0)
{
UINFO("Created %d/%d widgets...", i, findObject_->objects().size());
}
}
UINFO("Creating %d object widgets... done!", findObject_->objects().size());
ui_->actionSave_objects->setEnabled(true);
ui_->actionSave_session->setEnabled(true);
}
@ -1045,9 +1053,12 @@ void MainWindow::showObject(ObjWidget * obj)
ui_->verticalLayout_objects->insertLayout(ui_->verticalLayout_objects->count()-1, vLayout);
QByteArray ba;
if(obj->pixmap().width() > 0)
{
QBuffer buffer(&ba);
buffer.open(QIODevice::WriteOnly);
obj->pixmap().scaledToWidth(128).save(&buffer, "JPEG"); // writes image into JPEG format
}
imagesMap_.insert(obj->id(), ba);
// update objects size slider
@ -1411,6 +1422,14 @@ void MainWindow::update(const cv::Mat & image)
// add rectangle
QPen rectPen(obj->color());
rectPen.setWidth(Settings::getHomography_rectBorderWidth());
if(rect.isNull())
{
QMap<int, ObjSignature*>::const_iterator iter = findObject_->objects().constFind(id);
if(iter!=findObject_->objects().end())
{
rect = iter.value()->rect();
}
}
RectItem * rectItemScene = new RectItem(id, rect);
connect(rectItemScene, SIGNAL(hovered(int)), this, SLOT(rectHovered(int)));
rectItemScene->setPen(rectPen);

65
src/ObjSignature.h
View File

@ -34,6 +34,8 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include <QtCore/QRect>
#include <QtCore/QDataStream>
#include <QtCore/QByteArray>
#include <QtCore/QFileInfo>
#include <Compression.h>
namespace find_object {
@ -84,18 +86,46 @@ public:
keypoints_.at(j).size;
}
qint64 dataSize = descriptors_.elemSize()*descriptors_.cols*descriptors_.rows;
streamPtr << descriptors_.rows <<
descriptors_.cols <<
descriptors_.type() <<
dataSize;
streamPtr << QByteArray((char*)descriptors_.data, dataSize);
std::vector<unsigned char> bytes = compressData(descriptors_);
qint64 dataSize = bytes.size();
int old = 0;
if(dataSize <= std::numeric_limits<int>::max())
{
// old: rows, cols, type
streamPtr << old << old << old << dataSize;
streamPtr << QByteArray::fromRawData((const char*)bytes.data(), dataSize);
}
else
{
UERROR("Descriptors (compressed) are too large (%d MB) to be saved! Limit is 2 GB (based on max QByteArray size).",
dataSize/(1024*1024));
// old: rows, cols, type, dataSize
streamPtr << old << old << old << old;
streamPtr << QByteArray(); // empty
}
streamPtr << words_;
if(!image_.empty())
{
std::vector<unsigned char> bytes;
QString ext = QFileInfo(filePath_).suffix();
if(ext.isEmpty())
{
// default png
cv::imencode(".png", image_, bytes);
streamPtr << QByteArray((char*)bytes.data(), (int)bytes.size());
}
else
{
cv::imencode(std::string(".")+ext.toStdString(), image_, bytes);
}
streamPtr << QByteArray::fromRawData((const char*)bytes.data(), (int)bytes.size());
}
else
{
streamPtr << QByteArray();
}
streamPtr << rect_;
}
@ -120,15 +150,34 @@ public:
int rows,cols,type;
qint64 dataSize;
streamPtr >> rows >> cols >> type >> dataSize;
if(rows == 0 && cols == 0 && type == 0)
{
// compressed descriptors
UASSERT(dataSize <= std::numeric_limits<int>::max());
QByteArray data;
streamPtr >> data;
descriptors_ = uncompressData((unsigned const char*)data.data(), dataSize);
}
else
{
// old raw format
QByteArray data;
streamPtr >> data;
if(data.size())
{
descriptors_ = cv::Mat(rows, cols, type, data.data()).clone();
}
else if(dataSize)
{
UERROR("Error reading descriptor data for object=%d", id_);
}
}
streamPtr >> words_;
QByteArray image;
streamPtr >> image;
if(!ignoreImage)
if(!ignoreImage && image.size())
{
std::vector<unsigned char> bytes(image.size());
memcpy(bytes.data(), image.data(), image.size());

5
src/ObjWidget.cpp
View File

@ -265,9 +265,14 @@ void ObjWidget::setTextLabel(const QString & text)
}
void ObjWidget::updateImage(const QImage & image)
{
pixmap_ = QPixmap();
rect_ = QRect();
if(!image.isNull())
{
pixmap_ = QPixmap::fromImage(image);
rect_ = pixmap_.rect();
}
label_->setVisible(image.isNull());
}
void ObjWidget::updateData(const std::vector<cv::KeyPoint> & keypoints, const QMultiMap<int, int> & words)

57
src/Vocabulary.cpp
View File

@ -28,9 +28,11 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "find_object/Settings.h"
#include "find_object/utilite/ULogger.h"
#include "Compression.h"
#include "Vocabulary.h"
#include <QtCore/QVector>
#include <QDataStream>
#include <QTime>
#include <stdio.h>
#if CV_MAJOR_VERSION < 3
#include <opencv2/gpu/gpu.hpp>
@ -80,16 +82,31 @@ void Vocabulary::save(QDataStream & streamSessionPtr, bool saveVocabularyOnly) c
}
else
{
UINFO("Saving %d object references...", wordToObjects_.size());
streamSessionPtr << wordToObjects_;
}
// save words
qint64 dataSize = indexedDescriptors_.elemSize()*indexedDescriptors_.cols*indexedDescriptors_.rows;
streamSessionPtr << indexedDescriptors_.rows <<
indexedDescriptors_.cols <<
indexedDescriptors_.type() <<
dataSize;
streamSessionPtr << QByteArray((char*)indexedDescriptors_.data, dataSize);
qint64 rawDataSize = indexedDescriptors_.rows * indexedDescriptors_.cols * indexedDescriptors_.elemSize();
UINFO("Compressing words... (%dx%d, %d MB)", indexedDescriptors_.rows, indexedDescriptors_.cols, rawDataSize/(1024*1024));
std::vector<unsigned char> bytes = compressData(indexedDescriptors_);
qint64 dataSize = bytes.size();
UINFO("Compressed = %d MB", dataSize/(1024*1024));
int old = 0;
if(dataSize <= std::numeric_limits<int>::max())
{
// old: rows, cols, type
streamSessionPtr << old << old << old << dataSize;
streamSessionPtr << QByteArray::fromRawData((const char*)bytes.data(), dataSize);
}
else
{
UERROR("Vocabulary (compressed) is too large (%d MB) to be saved! Limit is 2 GB (based on max QByteArray size).",
dataSize/(1024*1024));
// old: rows, cols, type, dataSize
streamSessionPtr << old << old << old << old;
streamSessionPtr << QByteArray(); // empty
}
}
void Vocabulary::load(QDataStream & streamSessionPtr, bool loadVocabularyOnly)
@ -104,17 +121,45 @@ void Vocabulary::load(QDataStream & streamSessionPtr, bool loadVocabularyOnly)
}
else
{
UINFO("Loading words to objects references...");
streamSessionPtr >> wordToObjects_;
UINFO("Loaded %d object references...", wordToObjects_.size());
}
// load words
int rows,cols,type;
qint64 dataSize;
streamSessionPtr >> rows >> cols >> type >> dataSize;
if(rows == 0 && cols == 0 && type == 0)
{
// compressed vocabulary
UINFO("Loading words... (compressed format: %d MB)", dataSize/(1024*1024));
UASSERT(dataSize <= std::numeric_limits<int>::max());
QByteArray data;
streamSessionPtr >> data;
UINFO("Uncompress vocabulary...");
indexedDescriptors_ = uncompressData((unsigned const char*)data.data(), dataSize);
UINFO("Words: %dx%d (%d MB)", indexedDescriptors_.rows, indexedDescriptors_.cols,
(indexedDescriptors_.rows * indexedDescriptors_.cols * indexedDescriptors_.elemSize()) / (1024*1024));
}
else
{
// old raw format
UINFO("Loading words... (old format: %dx%d (%d MB))", rows, cols, dataSize/(1024*1024));
QByteArray data;
streamSessionPtr >> data;
UINFO("Allocate memory...");
if(data.size())
{
indexedDescriptors_ = cv::Mat(rows, cols, type, data.data()).clone();
}
else if(dataSize)
{
UERROR("Error reading vocabulary data...");
}
}
UINFO("Update vocabulary index...");
update();
}