def evaluateResults():
alfas = np.arange(0, 15)
rhos = np.arange(0, 1.1, 0.1)
TP = []
TN = []
FP = []
FN = []
precision = []
recall = []
F1 = []
F1_adaptative = []
bestF1 = 0
optAlpha = 0
# Compute the evaluation
'''
for alfa in alfas:
# One gaussian
dataset = "Traffic"
datasetGT = "TrafficGT"
colorSpace = 'gray' # 'gray', 'HSV', 'YCrCb', 'BGR'
gm.obtainGaussianModell(dataset, datasetGT, colorSpace, alfa)
TPi,TNi,FPi,FNi,precisioni,recalli,F1i = ev.evaluateFolder("./results/imagesGaussianModelling/","Traffic")
TP.append(TPi)
TN.append(TNi)
FP.append(FPi)
FN.append(FNi)
precision.append(precisioni)
recall.append(recalli)
F1.append(F1i)
# Adaptative gaussian
mu, sigma = ag.obtainGaussianModell("Traffic")
ag.foreground_substraction("Traffic", "TrafficGT", mu, sigma, alfa, 0) # rho equal to 0, in order to find the optimal alpha
aux,aux,aux,aux,aux,aux,F1i_adaptative = ev.evaluateFolder("./results/imagesAdaptativeGaussian/")
F1_adaptative.append(F1i_adaptative)
if F1i_adaptative > bestF1:
bestF1 = F1i_adaptative
optAlpha = alfa
print ('--- Alfa: ' + str(alfa) + ' --- Rho: 0')
print ('--- F1 Gaussian Model: ' + str(F1i))
print ('--- F1 Adaptative Gaussian Model: ' + str(F1i_adaptative))
'''
bestF1 = 0
optAlpha = 1.8
for rho in rhos:
# Adaptative gaussian
mu, sigma = ag.obtainGaussianModell("Highway")
ag.foreground_substraction("Highway", "HighwayGT", mu, sigma, optAlpha,
rho)
aux, aux, aux, aux, aux, aux, F1 = ev.evaluateFolder(
"./results/imagesAdaptativeGaussianModelling/")
if F1 > bestF1:
bestF1 = F1
optRho = rho
print('--- Rho: ' + str(rho) + ' --- ' + 'optimal Alpha: ' +
str(optAlpha))
print('--- F1 Adaptative Gaussian Model: ' + str(F1))
def evaluateTask1(colorSpace='gray'):
datasets = ['Fall'] # ["Highway","Fall","Traffic"]
datasetsGT = [el + "GT" for el in datasets]
alfaRange = np.arange(0, 5.001, 0.5)
rhoRange = np.arange(0, 1.1, 0.1)
if (os.path.exists('results/task1.json')):
with open('results/task1.json') as f:
datasetsResults = json.load(f)
print "Dataset results loaded"
else:
datasetsResults = dict()
for d, dGT in zip(datasets, datasetsGT):
if d not in datasetsResults.keys():
datasetsResults[d] = []
alfaResults = [el['Alfa'] for el in datasetsResults[d]]
for alfa in alfaRange:
if alfa in alfaResults:
print "Alfa " + str(
alfa) + " already done for this dataset. Skipping"
continue
print "Dataset " + d + ", alfa: " + str(alfa)
gm.obtainGaussianModell(d, dGT, colorSpace, alfa)
TPi, TNi, FPi, FNi, precisioni, recalli, F1i, _ = ev.evaluateFolder(
"./results/imagesGaussianModelling/", d)
performance = dict()
performance['Alfa'] = alfa
performance['TP'] = TPi
performance['TN'] = TNi
performance['FP'] = FPi
performance['FN'] = FNi
performance['precision'] = precisioni
performance['recall'] = recalli
performance['F1'] = F1i
print F1i
datasetsResults[d].append(performance)
with open('results/task1.json', "w") as f:
json.dump(datasetsResults, f)
alfaRange = sorted([el['Alfa'] for el in datasetsResults['Fall']])
fall = sorted(datasetsResults['Fall'], key=lambda k: k['Alfa'])
traffic = sorted(datasetsResults['Traffic'], key=lambda k: k['Alfa'])
highway = sorted(datasetsResults['Highway'], key=lambda k: k['Alfa'])
for el in fall:
print str(el['Alfa']) + ': ' + str(el['F1'])
fig = plt.figure()
plt.plot(alfaRange, [el['F1'] for el in fall], 'r', label='Fall')
plt.plot(alfaRange, [el['F1'] for el in traffic], 'g', label='Traffic')
plt.plot(alfaRange, [el['F1'] for el in highway], 'b', label='Highway')
plt.xlabel("Alfa")
plt.ylabel("F1")
plt.title("F1 vs alfa")
plt.axis([0, 5, 0, 1])
lgd = plt.legend(bbox_to_anchor=(1.3, 0.8))
plt.savefig("task1.png", bbox_extra_artists=(lgd, ), bbox_inches='tight')
plt.show()
#AUC:
y = np.array([el['precision'] for el in fall])
x = np.array([el['recall'] for el in fall])
AUC = trapz(y, x)
print " AUC of Fall: " + str(AUC)
y = np.array([el['precision'] for el in traffic])
x = np.array([el['recall'] for el in traffic])
AUC = trapz(y, x)
print " AUC of Traffic: " + str(AUC)
y = np.array([el['precision'] for el in highway])
x = np.array([el['recall'] for el in highway])
AUC = trapz(y, x)
print " AUC of Highway: " + str(AUC)
fig = plt.figure()
plt.plot([el['recall'] for el in fall], [el['precision'] for el in fall],
'r',
label='Fall')
plt.plot([el['recall'] for el in traffic],
[el['precision'] for el in traffic],
'g',
label='Traffic')
plt.plot([el['recall'] for el in highway],
[el['precision'] for el in highway],
'b',
label='Highway')
plt.axis([0, 1, 0, 1])
plt.title("Precision-Recall")
plt.xlabel("Recall")
plt.ylabel("Precision")
lgd = plt.legend(bbox_to_anchor=(1.3, 0.8))
plt.savefig("task1-curves.png",
bbox_extra_artists=(lgd, ),
bbox_inches='tight')
plt.show()
if not os.path.exists(gM_path):
os.makedirs(gM_path)
aG_path = results_path + "/imagesAdaptativeGaussianModelling/" #"/imagesAdaptativeGaussian/"
if not os.path.exists(aG_path):
os.makedirs(aG_path)
# Compute the evaluation
for alfa in alfas:
# One gaussian
dataset = "Highway"
datasetGT = "HighwayGT"
colorSpace = 'gray' # 'gray', 'HSV', 'YCrCb', 'BGR'
gm.obtainGaussianModell(dataset, datasetGT, colorSpace, alfa)
TPi, TNi, FPi, FNi, precisioni, recalli, F1i, auci = ev.evaluateFolder(
gM_path, datasetGT)
TP.append(TPi)
TN.append(TNi)
FP.append(FPi)
FN.append(FNi)
precision.append(precisioni)
recall.append(recalli)
F1.append(F1i)
auc.append(auci)
# Adaptative gaussian
mu, sigma = ag.obtainGaussianModell(dataset)
ag.foreground_substraction(
dataset, datasetGT, mu, sigma, alfa,
0) # rho equal to 0, in order to find the optimal alpha
alfaRange = np.concatenate((alfaPre, alfaRange, alfaPost), axis=0)
for alfa in alfaRange:
if alfa in alfaResults:
print "Alfa " + str(
alfa) + " already done for this dataset. Skipping"
continue
print "Dataset " + d + ", alfa: " + str(alfa)
if conf.isShadowremoval:
mu, sigma = acg.obtainGaussianModell(
d, conf.OptimalColorSpaces["ShadowRemoval"])
acg.foreground_substraction(
d, dGT, mu, sigma, alfa, conf.OptimalRhoParameter[d],
conf.OptimalColorSpaces["ShadowRemoval"])
TPi, TNi, FPi, FNi, precisioni, recalli, F1i = ev.evaluateFolder(
"./results/imagesAdaptativeGaussianModelling/", d)
else:
mu, sigma = acg.obtainGaussianModell(d, conf.OptimalColorSpaces[d])
acg.foreground_substraction(d, dGT, mu, sigma, alfa,
conf.OptimalRhoParameter[d],
conf.OptimalColorSpaces[d])
TPi, TNi, FPi, FNi, precisioni, recalli, F1i = ev.evaluateFolder(
"./results/imagesAdaptativeGaussianModelling/", d)
performance = dict()
performance['Alfa'] = alfa
performance['TP'] = TPi
performance['TN'] = TNi
performance['FP'] = FPi
performance['FN'] = FNi
performance['precision'] = precisioni
instance = np.stack([out, out, outError], axis=-1)
# cv2.imshow("OutputColor", instance * 255)
# cv2.imshow("Image", frame)
# cv2.imshow("Output", out * 255)
# cv2.imshow("GT", groundTruth*255)
# k = cv2.waitKey(5) & 0xff
# if k == 27:
# break
file_name = framesFiles[idx]
cv2.imwrite(
'./results/StaufferAndGrimson/' +
file_name[file_name.rfind('\\') + 1:], out)
videoOutput.write(instance * 255)
videoOutput.release()
if __name__ == "__main__":
dataset = "Highway"
datasetGT = "HighwayGT"
history = conf.history
nGauss = conf.nGauss
bgThresh = conf.bgThresh
StaufferAndGrimsonAlgorithm(dataset, datasetGT, history, nGauss, bgThresh)
aux, aux, aux, aux, aux, aux, F1 = ev.evaluateFolder(
"./results/StaufferAndGrimson/")
print('--- F1 Stauffer And Grimson Model: ' + str(F1))
aG_path = results_path + "/imagesAdaptativeGaussianModelling/" # "/imagesAdaptativeGaussian/"
if not os.path.exists(aG_path):
os.makedirs(aG_path)
if conf.isShadowremoval:
print('--- obtainGaussianModell')
mu, sigma = obtainGaussianModell(
dataset, conf.OptimalColorSpaces["ShadowRemoval"])
print('--- foreground_substraction')
foreground_substraction(dataset, datasetGT, mu, sigma,
conf.OptimalAlphaParameter[dataset],
conf.OptimalRhoParameter[dataset],
conf.OptimalColorSpaces["ShadowRemoval"])
print('--- evaluateFolder')
aux, aux, aux, aux, aux, aux, F1 = ev.evaluateFolder(aG_path, dataset)
else:
print('--- obtainGaussianModell')
mu, sigma = obtainGaussianModell(dataset,
conf.OptimalColorSpaces[dataset])
print('--- foreground_substraction')
foreground_substraction(dataset, datasetGT, mu, sigma,
conf.OptimalAlphaParameter[dataset],
conf.OptimalRhoParameter[dataset],
conf.OptimalColorSpaces[dataset])
print('--- evaluateFolder')
aux, aux, aux, aux, aux, aux, F1 = ev.evaluateFolder(aG_path, dataset)
print('--- DataSet: ' + dataset)
print('--- Rho: ' + str(conf.OptimalRhoParameter[dataset]) + ' --- ' +
' Alpha: ' + str(conf.OptimalAlphaParameter[dataset]))
Precision = []
Recall = []
F1 = []
for p in range(0, 200, 10):
print "Filtering objects smaller than " + str(p)
for f in files:
out = bwareaopen.bwareaopen(f, p)
if operativeSystem == 'posix':
#posix systems go here: ubuntu, debian, linux mint, red hat, etc, even osX (iew)
cv2.imwrite('results/areaFiltering/' + ID + f.split('/')[-1], out)
else:
#say hello to propietary software
cv2.imwrite('results/areaFiltering/' + ID + f.split('\\')[-1], out)
tp, tn, fp, fn, precision, recall, f1 = ev.evaluateFolder(
'results/areaFiltering/', ID)
for f in glob.glob('results/areaFiltering/*'):
os.remove(f)
print tp, tn, fp, fn, precision, recall, f1
TP.append(tp)
TN.append(tn)
FP.append(fp)
FN.append(fn)
Precision.append(precision)
Recall.append(recall)
F1.append(f1)
results['TP'] = TP
results['TN'] = TN
results['FP'] = FP
results['FN'] = FN