void Objectness::trainStageI()
{
vecM pX, nX;
pX.reserve(200000), nX.reserve(200000);
Mat xP1f, xN1f;
CV_Assert(matRead(_modelName + ".xP", xP1f) && matRead(_modelName + ".xN", xN1f));
for (int r = 0; r < xP1f.rows; r++)
pX.push_back(xP1f.row(r));
for (int r = 0; r < xN1f.rows; r++)
nX.push_back(xN1f.row(r));
Mat crntW = trainSVM(pX, nX, L1R_L2LOSS_SVC, 10, 1);
crntW = crntW.colRange(0, crntW.cols - 1).reshape(1, _W);
CV_Assert(crntW.size() == Size(_W, _W));
matWrite(_modelName + ".wS1", crntW);
}
void Objectness::generateTrianData()
{
const int NUM_TRAIN = _voc.trainNum;
const int FILTER_SZ = _W*_W;
vector<vector<Mat> > xTrainP(NUM_TRAIN), xTrainN(NUM_TRAIN);
vector<vecI> szTrainP(NUM_TRAIN); // Corresponding size index.
const int NUM_NEG_BOX = 100; // Number of negative windows sampled from each image
#pragma omp parallel for
for (int i = 0; i < NUM_TRAIN; i++) {
const int NUM_GT_BOX = (int)_voc.gtTrainBoxes[i].size();
vector<Mat> &xP = xTrainP[i], &xN = xTrainN[i];
vecI &szP = szTrainP[i];
xP.reserve(NUM_GT_BOX*4), szP.reserve(NUM_GT_BOX*4), xN.reserve(NUM_NEG_BOX);
Mat im3u = imread(format(_S(_voc.imgPathW), _S(_voc.trainSet[i])));
// Get positive training data
for (int k = 0; k < NUM_GT_BOX; k++) {
const Vec4i& bbgt = _voc.gtTrainBoxes[i][k];
vector<Vec4i> bbs; // bounding boxes;
vecI bbR; // Bounding box ratios
int nS = gtBndBoxSampling(bbgt, bbs, bbR);
for (int j = 0; j < nS; j++) {
bbs[j][2] = min(bbs[j][2], im3u.cols);
bbs[j][3] = min(bbs[j][3], im3u.rows);
Mat mag1f = getFeature(im3u, bbs[j]), magF1f;
flip(mag1f, magF1f, CV_FLIP_HORIZONTAL);
xP.push_back(mag1f);
xP.push_back(magF1f);
szP.push_back(bbR[j]);
szP.push_back(bbR[j]);
}
}
// Get negative training data
for (int k = 0; k < NUM_NEG_BOX; k++) {
int x1 = rand() % im3u.cols + 1, x2 = rand() % im3u.cols + 1;
int y1 = rand() % im3u.rows + 1, y2 = rand() % im3u.rows + 1;
Vec4i bb(min(x1, x2), min(y1, y2), max(x1, x2), max(y1, y2));
if (maxIntUnion(bb, _voc.gtTrainBoxes[i]) < 0.5)
xN.push_back(getFeature(im3u, bb));
}
}
const int NUM_R = _numT * _numT + 1;
vecI szCount(NUM_R); // Object counts of each size (combination of scale and aspect ratio)
int numP = 0, numN = 0, iP = 0, iN = 0;
for (int i = 0; i < NUM_TRAIN; i++) {
numP += xTrainP[i].size();
numN += xTrainN[i].size();
const vecI &rP = szTrainP[i];
for (size_t j = 0; j < rP.size(); j++)
szCount[rP[j]]++;
}
vecI szActive; // Indexes of active size
for (int r = 1; r < NUM_R; r++) {
if (szCount[r] > 50) // If only 50- positive samples at this size, ignore it.
szActive.push_back(r-1);
}
matWrite(_modelName + ".idx", Mat(szActive));
Mat xP1f(numP, FILTER_SZ, CV_32F), xN1f(numN, FILTER_SZ, CV_32F);
for (int i = 0; i < NUM_TRAIN; i++) {
vector<Mat> &xP = xTrainP[i], &xN = xTrainN[i];
for (size_t j = 0; j < xP.size(); j++)
memcpy(xP1f.ptr(iP++), xP[j].data, FILTER_SZ*sizeof(float));
for (size_t j = 0; j < xN.size(); j++)
memcpy(xN1f.ptr(iN++), xN[j].data, FILTER_SZ*sizeof(float));
}
CV_Assert(numP == iP && numN == iN);
matWrite(_modelName + ".xP", xP1f);
matWrite(_modelName + ".xN", xN1f);
}
void Objectness::trainStateII(int numPerSz)
{
loadTrainedModel();
const int NUM_TRAIN = _voc.trainNum;
vector<vecI> SZ(NUM_TRAIN), Y(NUM_TRAIN);
vector<vecF> VAL(NUM_TRAIN);
#pragma omp parallel for
for (int i = 0; i < _voc.trainNum; i++) {
const vector<Vec4i> &bbgts = _voc.gtTrainBoxes[i];
ValStructVec<float, Vec4i> valBoxes;
vecI &sz = SZ[i], &y = Y[i];
vecF &val = VAL[i];
CStr imgPath = format(_S(_voc.imgPathW), _S(_voc.trainSet[i]));
predictBBoxSI(imread(imgPath), valBoxes, sz, numPerSz, false);
const int num = valBoxes.size();
CV_Assert(sz.size() == num);
y.resize(num), val.resize(num);
for (int j = 0; j < num; j++) {
Vec4i bb = valBoxes[j];
val[j] = valBoxes(j);
y[j] = maxIntUnion(bb, bbgts) >= 0.5 ? 1 : -1;
}
}
const int NUM_SZ = _svmSzIdxs.size();
const int maxTrainNum = 100000;
vector<vecM> rXP(NUM_SZ), rXN(NUM_SZ);
for (int r = 0; r < NUM_SZ; r++) {
rXP[r].reserve(maxTrainNum);
rXN[r].reserve(1000000);
}
for (int i = 0; i < NUM_TRAIN; i++) {
const vecI &sz = SZ[i], &y = Y[i];
vecF &val = VAL[i];
int num = sz.size();
for (int j = 0; j < num; j++) {
int r = sz[j];
CV_Assert(r >= 0 && r < NUM_SZ);
if (y[j] == 1)
rXP[r].push_back(Mat(1, 1, CV_32F, &val[j]));
else
rXN[r].push_back(Mat(1, 1, CV_32F, &val[j]));
}
}
Mat wMat(NUM_SZ, 2, CV_32F);
for (int i = 0; i < NUM_SZ; i++) {
const vecM &xP = rXP[i], &xN = rXN[i];
if (xP.size() < 10 || xN.size() < 10)
printf("Warning %s:%d not enough training sample for r[%d] = %d. P = %d, N = %d\n", __FILE__, __LINE__, i, _svmSzIdxs[i], xP.size(), xN.size());
for (size_t k = 0; k < xP.size(); k++)
CV_Assert(xP[k].size() == Size(1, 1) && xP[k].type() == CV_32F);
Mat wr = trainSVM(xP, xN, L1R_L2LOSS_SVC, 100, 1);
CV_Assert(wr.size() == Size(2, 1));
wr.copyTo(wMat.row(i));
}
matWrite(_modelName + ".wS2", wMat);
_svmReW1f = wMat;
}
int main(int argc, char **argv)
{
char *inputFilename = NULL;
int inputFiles = 0;
char *matFilename = NULL;
busAssignment_t *busAssignment = busAssignment_create();
int bus = -1;
signalFormat_t signalFormat = signalFormat_Name;
measurement_t *measurement;
int ret = 1;
sint32 timeResolution = 10000;
parserFunction_t parserFunction = NULL;
program_name = argv[0];
/* parse arguments */
while (1) {
static struct option long_options[] = {
/* These options set a flag. */
{"verbose", no_argument, &verbose_flag, 1},
{"brief", no_argument, &verbose_flag, 0},
{"debug", no_argument, &debug_flag, 1},
/* These options don't set a flag.
We distinguish them by their indices. */
{"asc", required_argument, 0, 'a'},
{"bus", required_argument, 0, 'b'},
#ifdef HAVE_CLGREADER_H
{"clg", required_argument, 0, 'c'},
#endif
{"dbc", required_argument, 0, 'd'},
{"format", required_argument, 0, 'f'},
{"mat", required_argument, 0, 'm'},
{"timeres", required_argument, 0, 't'},
{"vsb", required_argument, 0, 'v'},
{"help", no_argument, NULL, 'h'},
{0, 0, 0, 0}
};
/* getopt_long stores the option index here. */
int option_index = 0;
int c;
c = getopt_long (argc, argv, "a:b:c:d:f:m:t:v:",
long_options, &option_index);
/* Detect the end of the options. */
if (c == -1) break;
switch (c) {
case 0:
break;
case 'a':
inputFilename = optarg;
parserFunction =ascReader_processFile;
inputFiles++;
break;
#ifdef HAVE_CLGREADER_H
case 'c':
inputFilename = optarg;
parserFunction =clgReader_processFile;
inputFiles++;
break;
#endif
case 'b':
bus = atoi(optarg);
break;
case 'd':
if(verbose_flag) {
if(bus == -1) {
fprintf(stderr, "Assigning DBC file %s to all busses\n", optarg);
} else {
fprintf(stderr, "Assigning DBC file %s to bus %d\n", optarg, bus);
}
}
busAssignment_associate(busAssignment, bus, optarg);
/* reset bus specification */
bus = -1;
break;
case 'm':
matFilename = optarg;
break;
case 'f':
if(!strcmp(optarg, "n")) {
signalFormat = signalFormat_Name;
} else if(!strcmp(optarg, "mn")) {
signalFormat = signalFormat_Message
| signalFormat_Name;
} else if(!strcmp(optarg, "dmn")) {
signalFormat = signalFormat_Database
| signalFormat_Message
| signalFormat_Name;
} else {
fprintf(stderr, "error: format must be 's', 'ms', or 'dms'\n");
usage_error();
}
matFilename = optarg;
break;
case 't':
timeResolution = atoi(optarg);
break;
case 'v':
inputFilename = optarg;
parserFunction = vsbReader_processFile;
inputFiles++;
break;
case 'h': help(); exit(0); break;
case '?':
/* getopt_long already printed an error message. */
usage_error();
break;
default:
fprintf(stderr, "error: unknown option %c\n", c);
busAssignment_free(busAssignment);
usage_error();
}
}
#ifdef YYDEBUG
if(debug_flag) {
extern int yydebug;
yydebug=1;
}
#endif
/* diagnose options */
if(inputFiles != 1) {
fprintf(stderr, "error: please specify exactly one input file\n");
busAssignment_free(busAssignment);
usage_error();
}
if(matFilename == NULL) {
fprintf(stderr, "error: MAT output filename not specified\n");
busAssignment_free(busAssignment);
usage_error();
}
/* parse DBC files */
busAssignment_parseDBC(busAssignment);
/* parse input file */
if(verbose_flag) {
if(inputFilename != NULL) {
fprintf(stderr,
"Parsing input file %s\n",
inputFilename?inputFilename:"<stdin>");
}
}
measurement = measurement_read(busAssignment,
inputFilename,
signalFormat,
timeResolution,
parserFunction);
if(measurement != NULL) {
/* write MAT file */
if(verbose_flag) {
fprintf(stderr, "Writing MAT file %s\n", matFilename);
}
matWrite(measurement, matFilename);
/* free memory */
measurement_free(measurement);
}
ret = 0;
usage_error:
busAssignment_free(busAssignment);
return ret;
}