void harriet_state::harriet(machine_config &config)
{
M68010(config, m_maincpu, 40_MHz_XTAL / 4); // MC68010FN10
m_maincpu->set_addrmap(AS_PROGRAM, &harriet_state::harriet_map);
MC68681(config, "duart", 3.6864_MHz_XTAL);
mc68901_device &mfp(MC68901(config, "mfp", 40_MHz_XTAL / 16));
mfp.set_timer_clock(2.4576_MHz_XTAL);
mfp.set_rx_clock(9600);
mfp.set_tx_clock(9600);
mfp.out_so_cb().set("rs232", FUNC(rs232_port_device::write_txd));
//mfp.out_tco_cb().set("mfp", FUNC(mc68901_device::rc_w));
//mfp.out_tdo_cb().set("mfp", FUNC(mc68901_device::tc_w));
HD63450(config, "dmac", 40_MHz_XTAL / 4, "maincpu"); // MC68450R10 (or HD68450Y-10)
M48T02(config, "timekpr");
NVRAM(config, "zpram", nvram_device::DEFAULT_ALL_0); // MK48Z02
rs232_port_device &rs232(RS232_PORT(config, "rs232", default_rs232_devices, "terminal"));
rs232.rxd_handler().set("mfp", FUNC(mc68901_device::write_rx));
rs232.rxd_handler().append("mfp", FUNC(mc68901_device::tbi_w));
NSCSI_BUS(config, "scsia");
NSCSI_CONNECTOR(config, "scsia:7").option_set("wdc", WD33C93A).clock(40_MHz_XTAL / 4);
//WD33C93(config, "wdcb", 40_MHz_XTAL / 4);
//IMSC012(config, "c012", 40_MHz_XTAL / 8); // INMOS IMSC012-P20S link adaptor
}
void threader::
call_end_handlers (
)
{
reg.m.lock();
const thread_id_type id = get_thread_id();
thread_id_type id_copy;
member_function_pointer<>::kernel_1a mfp;
// Remove all the member function pointers for this thread from the tree
// and call them.
while (reg.reg[id] != 0)
{
reg.reg.remove(id,id_copy,mfp);
reg.m.unlock();
mfp();
reg.m.lock();
}
reg.m.unlock();
}
void threader::
call_end_handlers (
)
{
reg.m.lock();
const thread_id_type id = get_thread_id();
thread_id_type id_copy;
unsigned long count = reg.reg.count(id);
member_function_pointer<>::kernel_1a mfp;
// Remove all the member function pointers for this thread from the tree
// and call them.
for (unsigned long i = 0; i < count; ++i)
{
reg.reg.remove(id,id_copy,mfp);
reg.m.unlock();
mfp();
reg.m.lock();
}
reg.m.unlock();
}
void genHardNegativesOnAnnotations(FeatGen* ldgFeatGen,const path& baseDir, datasets currDataset, bool writeOutWins, bool viz,string modelfile)
{
vector<path> negativePaths;
vector<path> negTrainPaths;
vector<path> negTestPaths;
FrameId firstTestFrame; // separates training and testing data
string learnFileStem;
vector<path> normPosWinPathsTrain; // if any use folder with normalized cropped windows
vector<path> normPosWinPathsTest; // if any use folder with normalized cropped windows
getLearnInfos(baseDir, currDataset, negativePaths, negTrainPaths, negTestPaths,normPosWinPathsTrain, normPosWinPathsTest, learnFileStem, firstTestFrame);
path negTrainFolder = baseDir / "learning" / "train_neg_hard_annot";
remove_all(negTrainFolder);
create_directories(negTrainFolder);
//std::ios_base::openmode mode = ;
//fs::ofstream trainFile((baseDir / "learning" / learnFileStem).string(), std::ios_base::out | ios_base::app);
//fs::ofstream trainFilesFile((baseDir / "learning" / learnFileStem).string() + ".files", std::ios_base::out | ios_base::app);
fs::ofstream trainFileHard((baseDir / "learning" / learnFileStem).string() + "_hard_annot");
if ( !(modelfile.length() > 0 ) )
{
modelfile = (baseDir / "learning" / learnFileStem).string() + ".model";
}
path mfp(modelfile);
if ( !exists(mfp))
{
cerr << "Modelfile does not exist: " << modelfile << endl;
exit(EXIT_FAILURE);
}
cout << "genHardNegativesOnAnnotations using model: " << modelfile << endl;
Detector pd(ldgFeatGen);
pd.setStartScale(1.0);
pd.setScaleStep(1.04);
pd.loadSVMWeights(modelfile);
float decThresh = 0;
int nww = ldgFeatGen->getWinW();
int nwh = ldgFeatGen->getWinH();
// use caltech annotations
FrameId startFrame;
startFrame.set= 0;
startFrame.video= 0;
startFrame.image= 0;
size_t featSize = getFeatLen(ldgFeatGen);
vector<path> annotFiles;
annotFiles = findAllAnnotations(baseDir,startFrame);
if ( annotFiles.size() == 0 )
{
cerr << "No annotations found." << endl;
exit(EXIT_FAILURE);
}
for ( vector<path>::iterator it (annotFiles.begin()); it != annotFiles.end(); ++it)
{
path vbbFile = *it;
path txtFolder = vbbFile.parent_path() / vbbFile.stem();
vector<path> v;
copy(directory_iterator(txtFolder),directory_iterator(),back_inserter(v));
sort(v.begin(),v.end());
int posTrainCounter = 0;
cout << "0% ";
cout.flush();
for ( vector<path>::iterator vit (v.begin()); vit != v.end(); ++vit)
{
path txtFile = *vit; // annotation txtFile
// corresponding image file
path imgFile = baseDir / "videos" / txtFile.parent_path().parent_path().stem() / txtFile.parent_path().stem() / txtFile.filename().stem();
bool isTrain = true;
if (frameLTEQ(firstTestFrame,parseFrameFromPath(imgFile)))
{
isTrain = false;
}
if ( skipPath(imgFile,startFrame) )
continue;
if ( !isTrain ) continue;
imgFile += findImageExtension(imgFile);
if ( !exists(imgFile) || !is_regular_file(imgFile))
{
cerr << "Could not find corresponding image file " <<imgFile <<endl;
cerr << "Export all .seq files using provided matlab code." << endl;
exit(EXIT_FAILURE);
}
// parse annotations from txtFile
fs::ifstream f(txtFile);
if (!f)
{
cerr << "cannot open file " << txtFile << endl;
exit(EXIT_FAILURE);
}
vector<Annotation> annots;
string buffer;
while (std::getline(f,buffer))
{
Annotation a = parseAnnotation(buffer);
if ( a.type == "person" )
{
//printAnnotation(a);
annots.push_back(a);
}
}
// extract normalized bb images
Mat img = imread(imgFile.string(), CV_LOAD_IMAGE_COLOR);
posTrainCounter++;
if ( (posTrainCounter % 100) == 0)
{
float donePerc = posTrainCounter / (float)v.size();
cout << floor(donePerc * 100) << "% ";
cout.flush();
}
path nf = imgFile;
string filePre = nf.parent_path().parent_path().stem().string() + nf.parent_path().stem().string() + "_" + nf.filename().stem().string();
vector< pair < Detection,double* > > detections;
vector<Detection> alldetections;
pd.detect(img,detections,alldetections,decThresh,true);
vector<bool> rectIsUsed;
rectIsUsed.resize(detections.size());
for ( size_t i = 0; i < detections.size(); i++)
{
pair < Detection,double* > mp = detections[i];
Detection det = mp.first;
Rect r = det.r;
double* f = mp.second;
// check if rectangle is annotated object
bool rectIsFree = true;
for (vector<Annotation>::iterator ait (annots.begin()); ait != annots.end(); ++ait)
{
Annotation an = *ait;
Rect a(an.x,an.y,an.width,an.height);
// calc intersection area
Rect inters = a & r;
if ( inters.area() > 0 )
{
// further analyze intersection
double ratio1 = (double) inters.area() / (double) r.area();
double ratio2 = (double) inters.area() / (double) a.area();
double ratio = min(ratio1,ratio2);
rectIsFree = !(ratio > 0.5);
if ( !rectIsFree )
{
break;
}
}
}
rectIsUsed[i] = rectIsFree;
if ( rectIsFree )
{
// save as negative example
writeFeatToSVMStream(f,trainFileHard,featSize,false);
}
else
{
// save as positive example
writeFeatToSVMStream(f,trainFileHard,featSize,true);
}
delete[] f;
}
if ( viz || writeOutWins )
{
Mat vizImg = img.clone();
if ( alldetections.size() < 100 )
{
for ( vector< Detection >::iterator it(alldetections.begin()); it != alldetections.end(); it++)
{
Detection det = *it;
Rect r = det.rNormWin;
//cout << r.x << " " << r.y << " " << r.width << "x" << r.height << " scale:" << det.scale << " decisionValue:" << det.decisionValue << endl;
rectangle(vizImg,r,Scalar(255,50,50),1,8);
}
}
for ( size_t i = 0; i < detections.size(); i++)
{
Detection det = detections[i].first;
Rect r = det.r;
if ( rectIsUsed[i])
{
rectangle(vizImg,r,Scalar(0,0,255),1,8);
}
else
{
rectangle(vizImg,r,Scalar(0,255,255),1,8);
}
}
displayAnnotations(vizImg,annots);
if ( writeOutWins )
{
string fileNameOnly = (filePre +"_hard.png");
path nwPath = negTrainFolder / fileNameOnly;
imwrite(nwPath.string(),vizImg);
}
if ( viz )
{
imshow("negative training image",vizImg);
waitKey(0);
destroyWindow("negative training image");
}
}
}
}
cout << "100% "<< endl;
trainFileHard.close();
}
void genHardNegatives(string retrainAppendFile, FeatGen* ldgFeatGen,const path& baseDir, datasets currDataset, bool writeOutWins, bool viz,string modelfile, double negBoundary)
{
bool debugOnlySomeWins = false;
vector<path> negativePaths;
vector<path> negTrainPaths;
vector<path> negTestPaths;
FrameId firstTestFrame; // separates training and testing data
string learnFileStem;
vector<path> normPosWinPathsTrain; // if any use folder with normalized cropped windows
vector<path> normPosWinPathsTest; // if any use folder with normalized cropped windows
getLearnInfos(baseDir, currDataset, negativePaths, negTrainPaths, negTestPaths,normPosWinPathsTrain, normPosWinPathsTest, learnFileStem, firstTestFrame);
path negTrainFolder = baseDir / "learning" / "train_neg_hard";
remove_all(negTrainFolder);
create_directories(negTrainFolder);
//std::ios_base::openmode mode = ;
//fs::ofstream trainFile((baseDir / "learning" / learnFileStem).string(), std::ios_base::out | ios_base::app);
//fs::ofstream trainFilesFile((baseDir / "learning" / learnFileStem).string() + ".files", std::ios_base::out | ios_base::app);
//fs::ofstream trainFileHard((baseDir / "learning" / learnFileStem).string() + "_hard");
fs::ofstream trainFileHard(retrainAppendFile,std::ios_base::out | ios_base::app);
if ( !(modelfile.length() > 0 ) )
{
modelfile = (baseDir / "learning" / learnFileStem).string() + ".model";
}
path mfp(modelfile);
if ( !exists(mfp))
{
cerr << "Modelfile does not exist: " << modelfile << endl;
exit(EXIT_FAILURE);
}
Detector pd(ldgFeatGen);
pd.setStartScale(1);
pd.setScaleStep(1.0718);
pd.loadSVMWeights(modelfile);
if ( negBoundary < 0 ) negBoundary *= -1.0;
float decThresh = -negBoundary;
int countedWindows = 0;
int nww = ldgFeatGen->getWinW();
int nwh = ldgFeatGen->getWinH();
cout << "genHardNegatives using model: " << mfp.string() << " dec: " << decThresh << endl;
for ( vector<path>::iterator pit(negTrainPaths.begin()); pit != negTrainPaths.end(); ++pit)
{
path np = *pit;
vector<path> v;
copy(directory_iterator(np),directory_iterator(),back_inserter(v));
sort(v.begin(),v.end());
size_t featSize = ldgFeatGen->getWindowFeatLen(); // feature size
int processedImages = 0;
cout << "0% ";
cout.flush();
for ( vector<path>::iterator it (v.begin()); it != v.end(); ++it)
{
if ( processedImages % 10 == 0 )
{
cout << floor((processedImages/(double) v.size()* 100.0 )) << "% ";
cout.flush();
}
if (debugOnlySomeWins && processedImages > 30)
{
break;
}
//if ( processedImages/(double) v.size()* 100.0 < 62.9 )
/*if ( processedImages < 770)
{
processedImages++;
continue;
}*/
path nf = *it;
string filePre = nf.parent_path().parent_path().stem().string() + nf.parent_path().stem().string() + "_" + nf.filename().stem().string();
Mat img = imread(nf.string(),CV_LOAD_IMAGE_COLOR);
vector< pair<Detection,double*> > detections;
vector<Detection> alldetections;
pd.detect(img,detections,alldetections,decThresh,true);
// pick random windows
vector<size_t> randPicks;
const int numPicks = 10;
if ( detections.size() > numPicks )
{
int picks = 0;
int dSize = detections.size();
while(picks != numPicks)
{
int randPick = rand() % dSize;
bool exists = false;
for ( int i = 0; i < randPicks.size(); i++ )
{
if ( randPicks[i] == randPick)
{
exists = true;
break;
}
}
if ( !exists )
{
picks++;
randPicks.push_back(randPick);
}
}
}
else
{
for ( int i = 0; i < detections.size(); i++ )
{
randPicks.push_back(i);
}
}
for ( size_t i = 0; i < randPicks.size(); i++)
{
pair < Detection,double* > mp = detections[randPicks[i]];
Detection det = mp.first;
Rect r = det.r;
double* f = mp.second;
writeFeatToSVMStream(f,trainFileHard,featSize,false);
countedWindows++;
}
// free cached features
for ( size_t i = 0; i < detections.size(); i++)
{
pair < Detection,double* > mp = detections[i];
double* f = mp.second;
delete [] f;
}
if ( viz || writeOutWins )
{
Mat vizImg = img.clone();
for ( vector< Detection >::iterator it(alldetections.begin()); it != alldetections.end(); it++)
{
Detection det = *it;
Rect r = det.rNormWin;
//cout << r.x << " " << r.y << " " << r.width << "x" << r.height << " scale:" << det.scale << " decisionValue:" << det.decisionValue << endl;
rectangle(vizImg,r,Scalar(255,50,50),1,8);
}
for ( size_t i = 0; i < detections.size(); i++)
{
Detection det = detections[i].first;
Rect r = det.r;
rectangle(vizImg,r,Scalar(0,0,255),1,8);
// render detection score
ostringstream detscoretxt;
detscoretxt.precision(3);
detscoretxt << det.decisionValue;
string text = detscoretxt.str();
int fontFace = FONT_HERSHEY_DUPLEX;
double fontScale = 0.5;
int thickness = 0;
int baseline=0;
Size textSize = getTextSize(text, fontFace,
fontScale, thickness, &baseline);
baseline += thickness;
Point textOrg(det.r.x + (det.r.width - textSize.width)/2.0,
det.r.y + (det.r.height - textSize.height)/2.0);
bool isPicked = false;
for ( size_t k = 0; k < randPicks.size(); k++ )
{
if ( randPicks[k] == i )
{
isPicked = true;
break;
}
}
if (isPicked )
{
rectangle(vizImg, textOrg + Point(0, baseline-3),
textOrg + Point(textSize.width, -textSize.height),
Scalar(80,200,80), CV_FILLED);
}
else
{
rectangle(vizImg, textOrg + Point(0, baseline-3),
textOrg + Point(textSize.width, -textSize.height),
Scalar(0,0,255), CV_FILLED);
}
// ... and the baseline first
//line(img, textOrg + Point(0, thickness),
// textOrg + Point(textSize.width, thickness),
// Scalar(0, 0, 255));
putText(vizImg, text, textOrg, fontFace, fontScale,
Scalar::all(255), thickness, 8);
}
if ( writeOutWins && detections.size() > 0 )
{
string fileNameOnly = (filePre +"_hard.png");
path nwPath = negTrainFolder / fileNameOnly;
imwrite(nwPath.string(),vizImg);
}
if ( viz )
{
imshow("negative training image",vizImg);
waitKey(0);
}
}
processedImages++;
}
cout << "100% " << endl;
}
//trainFile.close();
trainFileHard.close();
cout << countedWindows << " hard negatives found." << endl;
}