def mainLoop(modelType,modelArgs,positives,trueObjectBoxes,trainingList,featuresDir,featuresExt,modelOut,maxNegOverlap,iter):
pos,posIdx,ari,osi = positives
startTime = cu.tic()
if iter == 0:
## Random Negatives
print ' >>> RANDOM NEGATIVES'
neg,negIdx = learn.getRandomNegs(featuresDir,trainingList,featuresExt,pos.shape[1],maxVectorsCache,maxNegativeImages)
detectionsList = [ [x[0],'0.0']+x[1:]+['1'] for x in negIdx]
hards = {'features':np.zeros((0,neg.shape[1])),'index':[]}
lap = cu.toc('Random negatives matrix ('+str(neg.shape[0])+' instances)',startTime)
else:
## Mine hard negatives
print ' >>> MINING HARD NEGATIVES'
model = det.createDetector(modelType,modelArgs)
model.load(modelOut+'.'+ str( iter-1 ))
detectionsList = detector.detectObjects(model,trainingList,featuresDir,featuresExt,0.3,-10.0)
hards = cu.loadMatrixNoCompression(modelOut+'.hards').item()
lap = cu.toc('Hard negatives matrix ('+str(hards['features'].shape[0])+' instances)',startTime)
## Rank and clean negative detections
detectionsData = evaluation.loadDetections(detectionsList)
groundTruth = evaluation.loadGroundTruthAnnotations(trueObjectBoxes)
detectionsLog = evaluation.evaluateDetections(groundTruth,detectionsData,0.5,allowDuplicates=True) # overlapMeasure=validRegion,
evaluation.computePrecisionRecall(len(posIdx),detectionsLog['tp'],detectionsLog['fp'],'tmp.txt')
evaluation.computePrecAt(detectionsLog['tp'],[20,50,100,200,300,400,500])
logData = learn.parseRankedDetectionsFile(detectionsLog['log'],maxNegOverlap,maxNegativeVectors)
print ' >>> LOADING HARD NEGATIVES'
neg,negIdx = learn.loadHardNegativesFromList(featuresDir,logData['negExamples'],featuresExt,pos.shape[1],logData['negTaken'])
del(detectionsList,detectionsData,detectionsLog,logData)
lap = cu.toc('Ranked negatives matrix ('+str(neg.shape[0])+' instances)',lap)
neg = np.concatenate( (neg,hards['features']) )
negIdx = negIdx + hards['index']
## Learn Detector
clf = det.createDetector(modelType,modelArgs)
clf.learn(pos,neg,posIdx,negIdx)
clf.save(modelOut+'.'+str(iter))
lap = cu.toc('Classifier learned:',lap)
## Keep hard negatives for next iterations
scores = clf.predict(neg,negIdx)
hardNegsIdx = np.argsort(scores)
hardNeg = np.concatenate( (hards['features'], neg[hardNegsIdx[-cu.topHards:]]) )
negIdx = hards['index'] + [negIdx[j] for j in hardNegsIdx[-cu.topHards:]]
print 'Hard negatives:',hardNeg.shape[0]
hards = {'features':hardNeg, 'index':negIdx}
cu.saveMatrixNoCompression({'features':hardNeg,'index':negIdx},modelOut+'.hards')
print ' ** Iteration',iter,'done'
return {'detector':clf,'pos':pos,'posIdx':posIdx,'neg':neg,'negIdx':negIdx}
def buildDataSetWithTopDetections(detectionsData,topKPositives,featuresDir,featuresExt):
detections = evaluation.loadDetections(detectionsData)
m = cu.loadMatrix(featuresDir+'/'+detections[0][0]+'.'+featuresExt)
positivesInfo = {}
for i in range(topKPositives):
d = detections[i]
try:
positivesInfo[ d[0] ].append( map(int,d[2:6]) )
except:
positivesInfo[ d[0] ] = [ map(int,d[2:6]) ]
pos,posIdx = learn.loadHardNegativesFromList(featuresDir,positivesInfo,featuresExt,m.shape[1],topKPositives)
#ari,osi = learn.computeAspectRatioAndSizeIntervals(posIdx)
ari,osi = [],[]
print 'Positive Matrix with Top Detections ('+str(pos.shape[0])+' instances)'
return (pos,posIdx,ari,osi)
def buildDataSetWithTopDetections(detectionsData, topKPositives, featuresDir,
featuresExt):
detections = evaluation.loadDetections(detectionsData)
m = cu.loadMatrix(featuresDir + '/' + detections[0][0] + '.' + featuresExt)
positivesInfo = {}
for i in range(topKPositives):
d = detections[i]
try:
positivesInfo[d[0]].append(map(int, d[2:6]))
except:
positivesInfo[d[0]] = [map(int, d[2:6])]
pos, posIdx = learn.loadHardNegativesFromList(featuresDir, positivesInfo,
featuresExt, m.shape[1],
topKPositives)
#ari,osi = learn.computeAspectRatioAndSizeIntervals(posIdx)
ari, osi = [], []
print 'Positive Matrix with Top Detections (' + str(
pos.shape[0]) + ' instances)'
return (pos, posIdx, ari, osi)
try:
boxes = groundTruth[d[0]]
except:
continue
bestIoU = 0.0
for gt in boxes:
iou = det.IoU(d[2:6], gt)
if iou > bestIoU:
bestIoU = iou
print bestIoU
if bestIoU > minOverlap:
candidates.append(d)
return candidates
def saveCandidates(candidates, output):
out = open(output, 'w')
for k in candidates:
out.write(k[0] + ' ' + ' '.join(map(str, map(int, k[2:6]))) + '\n')
out.close()
if __name__ == "__main__":
params = cu.loadParams("detectionsFile groundTruths output")
detectionsData = [x.split() for x in open(params['detectionsFile'])]
detections = eval.loadDetections(detectionsData)
groundTruth = cu.loadBoxIndexFile(params['groundTruths'])
candidates = selectBestBoxes(detections, groundTruth, 0.5)
print 'Selected candidates:', len(candidates)
saveCandidates(candidates, params['output'])
def mainLoop(modelType, modelArgs, positives, trainingList, featuresDir,
featuresExt, modelOut, maxNegOverlap, iter):
pos, posIdx, ari, osi = positives
startTime = cu.tic()
if iter == 0:
## Random Negatives
print ' >>> RANDOM NEGATIVES'
neg, negIdx = learn.getRandomNegs(featuresDir, trainingList,
featuresExt, pos.shape[1],
maxVectorsCache, maxNegativeImages)
detectionsList = [[x[0], '0.0'] + x[1:] + ['1'] for x in negIdx]
hards = {'features': np.zeros((0, neg.shape[1])), 'index': []}
lap = cu.toc(
'Random negatives matrix (' + str(neg.shape[0]) + ' instances)',
startTime)
else:
## Mine hard negatives
print ' >>> MINING HARD NEGATIVES'
model = det.createDetector(modelType, modelArgs)
model.load(modelOut + '.' + str(iter - 1))
detectionsList = detector.detectObjects(
model, trainingList, featuresDir, featuresExt, 1.0, -10.0
) # For RCNN the overlap parameter is 0.3 not 1.0(no suppression)
hards = cu.loadMatrixNoCompression(modelOut + '.hards').item()
lap = cu.toc(
'Hard negatives matrix (' + str(hards['features'].shape[0]) +
' instances)', startTime)
## Rank and clean negative detections
detectionsData = evaluation.loadDetections(detectionsList)
groundTruth = evaluation.loadGroundTruthAnnotations(posIdx)
detectionsLog = evaluation.evaluateDetections(
groundTruth, detectionsData, 0.5,
allowDuplicates=False) #,overlapMeasure=det.overlap
evaluation.computePrecisionRecall(len(posIdx), detectionsLog['tp'],
detectionsLog['fp'], 'tmp.txt')
evaluation.computePrecAt(detectionsLog['tp'],
[20, 50, 100, 200, 300, 400, 500])
logData = learn.parseRankedDetectionsFile(detectionsLog['log'],
maxNegOverlap,
maxNegativeVectors)
print ' >>> LOADING HARD NEGATIVES'
neg, negIdx = learn.loadHardNegativesFromList(featuresDir,
logData['negExamples'],
featuresExt, pos.shape[1],
logData['negTaken'])
del (detectionsList, detectionsData, detectionsLog, logData)
lap = cu.toc(
'Ranked negatives matrix (' + str(neg.shape[0]) + ' instances)', lap)
neg = np.concatenate((neg, hards['features']))
negIdx = negIdx + hards['index']
## Learn Detector
clf = det.createDetector(modelType, modelArgs)
clf.learn(pos, neg, posIdx, negIdx)
clf.save(modelOut + '.' + str(iter))
lap = cu.toc('Classifier learned:', lap)
## Keep hard negatives for next iterations
scores = clf.predict(neg, negIdx)
hardNegsIdx = np.argsort(scores)
hardNeg = np.concatenate(
(hards['features'], neg[hardNegsIdx[-cu.topHards:]]))
negIdx = hards['index'] + [negIdx[j] for j in hardNegsIdx[-cu.topHards:]]
print 'Hard negatives:', hardNeg.shape[0]
hards = {'features': hardNeg, 'index': negIdx}
cu.saveMatrixNoCompression({
'features': hardNeg,
'index': negIdx
}, modelOut + '.hards')
print ' ** Iteration', iter, 'done'
return {
'detector': clf,
'pos': pos,
'posIdx': posIdx,
'neg': neg,
'negIdx': negIdx
}
def selectBestBoxes(detections, groundTruth, minOverlap):
candidates = []
for d in detections:
try: boxes = groundTruth[d[0]]
except: continue
bestIoU = 0.0
for gt in boxes:
iou = det.IoU(d[2:6],gt)
if iou > bestIoU:
bestIoU = iou
print bestIoU
if bestIoU > minOverlap:
candidates.append(d)
return candidates
def saveCandidates(candidates, output):
out = open(output, 'w')
for k in candidates:
out.write(k[0] + ' ' + ' '.join(map(str, map(int, k[2:6]) ) ) + '\n')
out.close()
if __name__ == "__main__":
params = cu.loadParams("detectionsFile groundTruths output")
detectionsData = [x.split() for x in open(params['detectionsFile'])]
detections = eval.loadDetections(detectionsData)
groundTruth = cu.loadBoxIndexFile(params['groundTruths'])
candidates = selectBestBoxes(detections, groundTruth, 0.5)
print 'Selected candidates:', len(candidates)
saveCandidates(candidates, params['output'])
