frame = cv2.imread(from_img_path + picName)
numFrame += 1
handcls, handbox, handscore = demo(net, frame)
nhand = handbox.shape[0]
nTP = 0
if nhand > 0:
# cv2.imwrite(toDir+picName, frame)
writeFile.write(picName + ' ' + str(nhand))
for i in xrange(0, nhand):
writeFile.write(' hand %d %d %d %d' %
(handbox[i][0], handbox[i][1],
handbox[i][2], handbox[i][3]))
if (n_hand_gt) == 0:
n_false_re += 1
elif (IOU_multi(handbox[i], box_gt) < iouN):
n_false_re += 1
writeFile.write("\n")
if numFrame % 100 == 0:
print numFrame, "faults=%d" % (n_false_re)
print "model%3dw: %6d faults in %6d frame" % (model_idx, n_false_re,
numFrame)
totalResultsFile.write("\nmodel%3dw: %6d faults in %6d frame" %
(model_idx, n_false_re, numFrame))
negSet = np.append(negSet, n_false_re)
writeFile.close()
print negSet
print np.where(negSet == min(negSet))
print "best model %d negs, at %d" % (np.min(negSet),
'.xml')
n_pos_gt += nhand_gt
frame = cv2.imread(from_img_path + picName)
numFrame += 1
handcls, handbox, handscore = demo(net, frame)
nhand = handbox.shape[0]
n_pos_re += nhand
nTP = 0
if nhand > 0:
writeFile.write(picName + ' ' + str(nhand))
for i in xrange(0, nhand):
writeFile.write(' hand %d %d %d %d' %
(handbox[i][0], handbox[i][1],
handbox[i][2], handbox[i][3]))
if (IOU_multi(handbox[i], box_gt) > iouP):
nTP += 1
writeFile.write("\n")
n_pos_tp += nTP
modelrecall = float(n_pos_tp) / float(n_pos_gt)
recallSet = np.append(recallSet, modelrecall)
modelprec = float(n_pos_tp) / float(n_pos_re)
precSet = np.append(precSet, modelprec)
print "model%3dw: gt =%4d, re =%4d, tp =%4d, prec=%.4f, recall=%.4f"%(model_idx, n_pos_gt,n_pos_re, n_pos_tp, \
float(n_pos_tp) / float(n_pos_re), float(n_pos_tp) / float(n_pos_gt))
totalResultsFile.write( "\nmodel%3dw: gt =%4d, re =%4d, tp =%4d, prec=%.4f, recall=%.4f"%(model_idx, n_pos_gt,n_pos_re, n_pos_tp, \
float(n_pos_tp) / float(n_pos_re), float(n_pos_tp) / float(n_pos_gt)) )
writeFile.close()
print np.min(precSet), np.where(precSet == min(precSet))
print np.max(recallSet), np.where(recallSet == max(recallSet))
print np.mean(recallSet)
for gt_idx in xrange(box_gt.shape[0]):
if min(box_gt[gt_idx][3] - box_gt[gt_idx][1],
box_gt[gt_idx][2] - box_gt[gt_idx][0]) < min(
frame.shape[0], frame.shape[1]) / 6.:
# print box_gt[gt_idx], frame.shape
small_flag = True
if (small_flag): #small box in gt: no need to recall
write_small.write(fileame + "\n")
continue
copyframe = frame.copy()
numFrame += 1
handcls, handbox, handscore = demo(net, frame, CONF_THRESH, NMS_THRESH)
nhand = handbox.shape[0]
nTP = 0
if nhand > 0: #recalled: find iou unright
for i in xrange(0, nhand):
iou = IOU_multi(handbox[i], box_gt)
if (iou <= iouM and iou > iouS):
cv2.imwrite(toDir + "_S/" + fileame + '.jpg', frame)
write_iouS.write(fileame + "\n")
elif (iou <= iouB and iou > iouM):
cv2.imwrite(toDir + "_B/" + fileame + '.jpg', frame)
write_iouB.write(fileame + "\n")
else: #unrecalled: savelists
write_unrecall.write(fileame + "\n")
cv2.imwrite(toDir + "_U/" + fileame + '.jpg', frame)
if numFrame % 100 == 0:
print numFrame
print "%6d negInNeg and %6d negInNull in %6d frame" % (
n_false_neg, n_false_null, numFrame)