def train(ds):
C = 2.0
gamma = 2.0
svm = cv2.ml.SVM_create()
svm.setType(cv2.ml.SVM_C_SVC)
svm.setKernel(cv2.ml.SVM_RBF)
svm.setC(C)
svm.setGamma(gamma)
trainData = np.zeros([ds.frameCount, 3*3*4*9])
trainLabels = np.zeros([ds.frameCount, 1])
i = 0
while (ds.getNextFrame()):
dets = Detection(ds.im)
h, w, _ = np.shape(ds.im)
hogs = dets.makeHogs(0, h, 0, w, 4, 4)
util.plotHelper(dets.visualHogs(0, h, 0, w, hogs), "bla", True)
v = dets.blockNormalization(hogs, 2, 2)
trainData[i] = v
if ds.getFilename()[0:3] == "pos":
trainLabels[i,0] = 1
elif ds.getFilename()[0:3] == "neg":
trainLabels[i,0] = 0
else:
raise Exception("Picture is not annotated !")
i = i + 1
#print(trainLabels)
svm.train(np.float32(trainData), cv2.ml.ROW_SAMPLE, np.int32(trainLabels))
result = svm.predict(np.float32(trainData))[1]
svm.save("model.dat");
#print("Result:", result)
print("GT Predict Name")
for i in range(ds.frameCount):
print(trainLabels[i,0], " ", result[i, 0], " ", ds.getFilename(i) )
return svm
def abcdefg():
ds = Dataset.Dataset("../data/circles/train/many/", "jpg")
ds.getNextFrame()
maxH, maxW, _ = np.shape(ds.im)
while ds.getNextFrame():
h, w, _ = np.shape(ds.im)
maxH = max(h, maxH)
maxW = max(w, maxW)
ds.aktuFrameIndex = -1
#scale all to height max-height:
ims = []
while ds.getNextFrame():
if not "pos" in ds.getFilename():
continue
h = np.shape(ds.im)[0]
f = maxH / h
im = cv2.resize(ds.im, (0, 0),
fx=f,
fy=f,
interpolation=cv2.INTER_LINEAR)
h, w, _ = np.shape(im)
print(h, w, maxH, maxW)
if w < maxW:
np.hstack((im, np.zeros([maxH, maxW - w, 3])))
im, _ = util.magAndAngle(im)
ims.append(im)
#util.plotHelper(im)
imRes = np.zeros([maxH, maxW])
for im in ims:
imRes = imRes + im
imRes = np.log10(1 + imRes)
util.plotHelper(imRes)
def testProjPic():
kam = kamera("c:/here_are_the_frames/calibration/iscam2.CALI", "")
dstIm = cv2.imread("c:/here_are_the_frames/00000002.jpg")
srcIm = cv2.imread("C:/here_are_the_frames/test2/123.jpg")
fp = np.matrix([0, -1, 0]).T
kam.projectPicture(fp, 0.7, 1.80, srcIm, dstIm)
util.plotHelper(dstIm, "Ziel", False)
util.plotHelper(srcIm, "Source")
def makeHogs(self, startY, height, startX, width, numOfCellsY, numOfCellsX): self.displayBild = self.mag.copy() self.startY = startY self.height = height self.startX = startX self.width = width deltaY = height / numOfCellsY #height of one cell deltaX = width / numOfCellsX #width of one cell hogs = np.zeros([numOfCellsY, numOfCellsX, 9]) for y in range(numOfCellsY): sY = startY + y * deltaY eY = round(sY + deltaY) sY = round(sY) for x in range(numOfCellsX): sX = startX + x * deltaX eX = round(sX + deltaX) sX = round(sX) hogs[y,x] = self.makeHog(sY, eY, sX, eX) util.plotHelper(self.displayBild, warte=False) return hogs
def makeManyTrainPics(srcPath, dstPath):
pos = 0
neg = 0
dataS = Dataset.Dataset(srcPath, "jpg")
while dataS.getNextFrame():
util.plotHelper(dataS.im, "bla")
#if "pos" in dataS.getFilename():
if "pos" in dataS.getFilename():
for angle in range(0, 360, 20):
im2 = util.rotatePicture(dataS.im, angle, True,
(255, 255, 255))
util.plotHelper(im2, "rotated")
cv2.imwrite(dstPath + "pos_" + "{:>04d}".format(pos) + ".jpg",
im2)
pos = pos + 1
elif "neg" in dataS.getFilename():
for angle in range(0, 360, 20):
im2 = util.rotatePicture(dataS.im, angle, True,
(255, 255, 255))
util.plotHelper(im2, "rotated")
cv2.imwrite(dstPath + "neg_" + "{:>04d}".format(neg) + ".jpg",
im2)
neg = neg + 1
else:
raise RuntimeError("File " + str(dataS.getFilename()) +
" is neither pos nor neg !?")
def projectTestsOnFrame():
#srcPath = "../data/circles/test/"
#dstPath = "../data/circles/test/many/"
#for d in glob.glob(dstPath + "*.jpg"):
# print ("Loesche: " + d)
# os.remove(d)
#makeManyTrainPics(srcPath, dstPath)
cam = kamera.kamera("../data/calibration/iscam2.cali", "")
dstPath = "C:/here_are_the_frames/00000002.jpg"
dstIm = cv2.imread(dstPath)
testCount = 5
srcPath = "../data/circles/test/many/"
ds = Dataset.Dataset(srcPath, "jpg")
r = np.random.randint(0, ds.frameCount, testCount)
dstH, dstW, _ = np.shape(dstIm)
for i in range(testCount):
srcIm = ds.getFrame(r[i])
util.plotHelper(srcIm, "bla_" + str(i), False)
fpUV = np.matrix([(dstW - 100) * np.random.rand() + 50,
(dstH - 100) * np.random.rand() + 50]).T
fp = cam.unproj(fpUV)
cam.projectPicture(fp, 1, 1, srcIm, dstIm)
util.plotHelper(dstIm)
def predict(ds, svm, maske):
testData = np.zeros([ds.frameCount, 3*3*4*9])
gtLabels = np.zeros([ds.frameCount, 1])
i = 0
while (ds.getNextFrame()):
det = Detection(ds.im, maske)
h, w, _ = np.shape(ds.im)
hogs = det.makeHogs(0, h, 0, w, 4, 4)
util.plotHelper(det.visualHogs(0, h, 0, w, hogs), "bla", True)
v = det.blockNormalization(hogs, 2, 2)
testData[i] = v
if ds.getFilename()[0:3] == "pos":
gtLabels[i,0] = 1
elif ds.getFilename()[0:3] == "neg":
gtLabels[i,0] = 0
else:
print("Fehler")
i = i + 1
result = svm.predict(np.float32(testData))[1]
print("GT Predict Name")
for i in range(ds.frameCount):
print(gtLabels[i,0], " ", result[i, 0], " ", ds.getFilename(i) )
def slide(self):
imRes = im.copy()
for y in range(0, np.shape(im)[0], 4):
for x in range(0, np.shape(im)[1], 4):
dstIm = im.copy()
fpUV = np.matrix([x, y]).T
fp = self.cam.unproj(fpUV)
upV = np.matrix([(0, ), (0, ), (1, )])
rightV = self.cam.kamkoord2weltkoord(
np.matrix([(1, ), (0, ),
(0, )])) - self.cam.kamkoord2weltkoord(
np.matrix([(0, ), (0, ), (0, )]))
#rightV = np.matrix( [(1,), (0,), (0,)] )
#print("Norm von rightV", np.linalg.norm(rightV))
rightV = (1.0 / np.linalg.norm(rightV)) * rightV
#print("rightV: ", rightV)
v1 = fp - (self.width / 2.0) * rightV
v2 = fp + (self.width / 2.0) * rightV
v3 = v1 + self.height * upV
v4 = v2 + self.height * upV
imaP1 = self.cam.proj(v1).T #row-vectors
imaP2 = self.cam.proj(v2).T
imaP3 = self.cam.proj(v3).T
imaP4 = self.cam.proj(v4).T
imaPFP = self.cam.proj(fp).T
imaPHP = self.cam.proj(fp + upV).T
util.plotLine(dstIm, imaPHP, imaPFP)
util.plotLine(dstIm, imaP1, imaP2)
util.plotCircle(dstIm, 3, (255, 255, 0), imaP1, imaP2)
x1, y1, x2, y2 = util.boundingRect(imaP1, imaP2, imaP3, imaP4)
util.plotCircle(dstIm, 3, (255, 255, 0), imaP1, imaP2, imaP3,
imaP4)
util.plotCircle(dstIm, 5, (255, 255, 255), imaPFP, imaPHP)
util.plotRect(dstIm, x1, y1, x2, y2)
x1 = int(x1)
x2 = int(x2)
y1 = int(y1)
y2 = int(y2)
imaP1 = np.squeeze(np.asarray(imaP1))
imaP2 = np.squeeze(np.asarray(imaP2))
imaP3 = np.squeeze(np.asarray(imaP3))
imaP4 = np.squeeze(np.asarray(imaP4))
dstH, dstW, _ = np.shape(dstIm)
y1 = util.clamp(y1, 0, dstH - 1)
y2 = util.clamp(y2, 0, dstH - 1)
x1 = util.clamp(x1, 0, dstW - 1)
x2 = util.clamp(x2, 0, dstW - 1)
if y2 > y1 + 8 and x2 > x1 + 8:
testData = np.zeros([1, 3 * 3 * 4 * 9])
h = y2 - y1 + 1
w = x2 - x1 + 1
hogs = self.det.makeHogs(y1, h, x1, w, 4, 4)
#util.plotHelper(self.det.visualHogs(y1, h, x1, w, hogs), "bla", False)
v = self.det.blockNormalization(hogs, 2, 2)
testData[0] = v
result = self.svm.predict(np.float32(testData))[1]
#print("GT Predict Name")
#for i in range(ds.frameCount):
# print(gtLabels[i,0], " ", result[i, 0], " ", ds.getFilename(i) )
#print(result[0, 0])
if result[0, 0]:
util.plotRect(imRes, x1, y1, x2, y2, (255, 0, 0))
print("!!!!!!!!!!!!!!!!!!!!!!!!!")
#util.plotRect(dstIm, x1, y1, x2, y2, (255,0,0))
#util.plotHelper(dstIm, "In Bearbeitung", False)
print(y)
util.plotHelper(imRes, "Im Res")