def Recall(imageList,model,sample_number=5000,thresh=0.3):
correct = 0
obj_num = 0
count = 0
for image in imageList:
count += 1
#Get prediction from neural network
img = crop_detection(image.imgPath,new_width=448,new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
out = out[0]
#for each ground truth, see we have predicted a corresponding result
for i in range(49):
preds = out[i*25:i*25+25]
row = i/7
col = i%7
box = image.boxes[row][col]
if(box.has_obj):
for obj in box.objs:
obj_num += 1
true_class = obj.class_num
#see what we predict
if(preds[24] > thresh):
predcit_class = np.argmax(preds[4:24])
if(predcit_class == true_class):
correct += 1
return correct*1.0/obj_num
def Recall(imageList, model, sample_number=5000, thresh=0.3):
correct = 0
obj_num = 0
count = 0
for image in imageList:
count += 1
#Get prediction from neural network
img = crop_detection(image.imgPath, new_width=448, new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
out = out[0]
#for each ground truth, see we have predicted a corresponding result
for i in range(49):
preds = out[i * 25:i * 25 + 25]
row = i / 7
col = i % 7
box = image.boxes[row][col]
if (box.has_obj):
for obj in box.objs:
obj_num += 1
true_class = obj.class_num
#see what we predict
if (preds[24] > thresh):
predcit_class = np.argmax(preds[4:24])
if (predcit_class == true_class):
correct += 1
return correct * 1.0 / obj_num
def Acc(imageList,model,sample_number=5000,thresh=0.3):
correct = 0
object_num = 0
count = 0
for image in imageList:
count += 1
if(count % 500 == 0):
print 'Image number:', count
#Get prediction from neural network
img = crop_detection(image.imgPath,new_width=448,new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
out = out[0]
for i in range(49):
preds = out[i*25:(i+1)*25]
if(preds[24] > thresh):
object_num += 1
row = i/7
col = i%7
'''
centerx = 64 * col + 64 * preds[0]
centery = 64 * row + 64 * preds[1]
h = preds[2] * preds[2]
h = h * 448.0
w = preds[3] * preds[3]
w = w * 448.0
left = centerx - w/2.0
right = centerx + w/2.0
up = centery - h/2.0
down = centery + h/2.0
if(left < 0): left = 0
if(right > 448): right = 447
if(up < 0): up = 0
if(down > 448): down = 447
'''
class_num = np.argmax(preds[4:24])
#Ground Truth
box = image.boxes[row][col]
if(box.has_obj):
for obj in box.objs:
true_class = obj.class_num
if(true_class == class_num):
correct += 1
break
return correct*1.0/object_num
def Acc(imageList, model, sample_number=5000, thresh=0.3):
correct = 0
object_num = 0
count = 0
for image in imageList:
count += 1
if (count % 500 == 0):
print 'Image number:', count
#Get prediction from neural network
img = crop_detection(image.imgPath, new_width=448, new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
predictions = out[0]
#Post process predicting results
boxes = convert_yolo_detections(predictions)
boxes = do_nms_sort(boxes)
for i in range(len(boxes)):
preds = boxes[i].probs
not_all_zeros = np.any(preds)
if (not_all_zeros):
object_num += 1
x = boxes[i].x
y = boxes[i].y
h = boxes[i].h
w = boxes[i].w
class_num = boxes[i].class_num
row = boxes[i].row
col = boxes[i].col
#Ground Truth
box = image.boxes[row][col]
if (box.has_obj):
for obj in box.objs:
true_class = obj.class_num
if (true_class == class_num):
correct += 1
break
return correct * 1.0 / object_num
def Acc(imageList,model,sample_number=5000,thresh=0.3):
correct = 0
object_num = 0
count = 0
for image in imageList:
count += 1
if(count % 500 == 0):
print 'Image number:', count
#Get prediction from neural network
img = crop_detection(image.imgPath,new_width=448,new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
predictions = out[0]
#Post process predicting results
boxes = convert_yolo_detections(predictions)
boxes = do_nms_sort(boxes)
for i in range(len(boxes)):
preds = boxes[i].probs
not_all_zeros = np.any(preds)
if(not_all_zeros):
object_num += 1
x = boxes[i].x
y = boxes[i].y
h = boxes[i].h
w = boxes[i].w
class_num = boxes[i].class_num
row = boxes[i].row
col = boxes[i].col
#Ground Truth
box = image.boxes[row][col]
if(box.has_obj):
for obj in box.objs:
true_class = obj.class_num
if(true_class == class_num):
correct += 1
break
return correct*1.0/object_num
def Acc(imageList,model,sample_number=5000,thresh=0.3):
fcat = open('comp4_det_test_cat.txt','write')
fcar = open('comp4_det_test_car.txt','write')
fbottle = open('comp4_det_test_bottle.txt','write')
ftrain = open('comp4_det_test_train.txt','write')
count = 0
for image in imageList:
img = Image.open(image.imgPath)
img_name = image.imgPath.split('/')[-1].split('.')[0]
W,H = img.size
drawable = ImageDraw.Draw(img)
count += 1
if(count % 500 == 0):
print 'Image number:', count
#Get prediction from neural network
imgg = crop_detection(image.imgPath,new_width=448,new_height=448)
imgg = np.expand_dims(imgg, axis=0)
out = model.predict(imgg)
predictions = out[0]
#Post process predicting results
boxes = convert_yolo_detections(predictions)
boxes = do_nms_sort(boxes)
for i in range(len(boxes)):
box = boxes[i]
prob = np.max(box.probs)
class_num = boxes[i].class_num
if((class_num in [4,6,7,18]) and prob >= 0.2):#if this box contains a car
centerx = boxes[i].x
centery = boxes[i].y
h = boxes[i].h
w = boxes[i].w
left = centerx - w/2.0
right = centerx + w/2.0
up = centery - h/2.0
down = centery + h/2.0
if(left < 0): left = 0
if(right > 448): right = 447
if(up < 0): up = 0
if(down > 448): down = 447
#convert the coords to image before streching to 448*448
left = left/448.0 * W
right = right/448.0 * W
up = up/448.0 * H
down = down/448.0 * H
#drawable.rectangle([left,up,right,down],outline="red")
if(class_num == 4):
fbottle.write(img_name+' '+str(round(prob,6))+' '+str(round(left,6))+' '+str(round(up,6))+' '+str(round(right,6))+' '+str(round(down,6))+'\n')
elif(class_num == 6):
fcar.write(img_name+' '+str(round(prob,6))+' '+str(round(left,6))+' '+str(round(up,6))+' '+str(round(right,6))+' '+str(round(down,6))+'\n')
elif(class_num == 7):
fcat.write(img_name+' '+str(round(prob,6))+' '+str(round(left,6))+' '+str(round(up,6))+' '+str(round(right,6))+' '+str(round(down,6))+'\n')
elif(class_num == 18):
ftrain.write(img_name+' '+str(round(prob,6))+' '+str(round(left,6))+' '+str(round(up,6))+' '+str(round(right,6))+' '+str(round(down,6))+'\n')
#img.save(os.path.join(os.getcwd(),'results',img_name+'.jpg'))
return
"pottedplant", "sheep", "sofa", "train", "tvmonitor"
]
#Predict output
#image = readImg(os.path.join(os.getcwd(),'Yolo_dog.img'),h=448,w=448)
images_path = '/home/media/Documents/YOLO.keras/images'
img_names = []
for root, dirs, files in os.walk(images_path):
for name in files:
img_names.append('images/' + name)
#img_names = ['images/eagle.jpg','images/000047.jpg','images/000009.jpg']
for img_name in img_names:
print img_name
img = crop_detection(os.path.join(os.getcwd(), img_name),
new_width=448,
new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
predictions = out[0]
#Post process predicting results
boxes = convert_yolo_detections(predictions)
boxes = do_nms_sort(boxes)
#print len(boxes)
draw_detections(
os.path.join('/home/media/Documents/YOLO.keras/', img_name), 0.2,
boxes, 20, labels,
img_name.split('/')[1])
print
return loss
model = model_from_json(open('Tiny_Yolo_Architecture.json').read(),custom_objects={'custom_loss':custom_loss})
model.load_weights('weights2.hdf5')
#Predict output
#image = readImg(os.path.join(os.getcwd(),'Yolo_dog.img'),h=448,w=448)
images_path = '/home/media/Documents/YOLO.keras/images'
img_names = []
for root, dirs, files in os.walk(images_path):
for name in files:
img_names.append('images/'+name)
#img_names = ['images/eagle.jpg','images/000047.jpg','images/000009.jpg']
for img_name in img_names:
print img_name
img,im = crop_detection(os.path.join(os.getcwd(),img_name),new_width=448,new_height=448,save=True)
im.save(os.getcwd()+'/images/'+img_name.split('/')[1].split('.')[0]+'_resize.jpg')
img = np.expand_dims(img, axis=0)
out = model.predict(img)
out = out[0]
#Post process predicting results
thresh = 0.3
imgPath = os.path.join(os.getcwd(),img_name)
img = Image.open(imgPath)
img_draw = Image.open(os.getcwd()+'/images/'+img_name.split('/')[1].split('.')[0]+'_resize.jpg')
drawable = ImageDraw.Draw(img_draw)
labels = ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train","tvmonitor"]
for i in range(49):
preds = out[i*25:(i+1)*25]
if(preds[24] > thresh):
def Acc(imageList, model, sample_number=5000, thresh=0.3):
classes = 20
side = 7
count = 0
ap = 0.0
rec = 0.0
image_num = 0
for image in imageList:
count += 1
if (count % 500 == 0):
print 'Image number:', count
#Get prediction from neural network
img = crop_detection(image.imgPath, new_width=448, new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
predictions = out[0]
#Post process predicting results
boxes = convert_yolo_detections(predictions)
boxes = do_nms_sort(boxes)
#pack ground truth boxes to true_boxes
truth_boxes = []
for i in range(side):
for j in range(side):
box_i_j = image.boxes[i][j]
if (box_i_j.has_obj):
for obj in box_i_j.objs:
new_box = box(classes)
new_box.row = i
new_box.col = j
new_box.x = (j + obj.x) * 64
new_box.y = (i + obj.y) * 64
new_box.h = (obj.h**2) * 448
new_box.w = (obj.w**2) * 448
new_box.class_num = obj.class_num
truth_boxes.append(new_box)
#assign detections to ground truth
n_pos = len(truth_boxes)
tp = np.zeros(len(boxes))
gt_covered = np.zeros(n_pos)
for i in range(len(boxes)):
#find the one ground truth with highest iou
boxa = boxes[i]
ovmax = -1
max_index = 0
for j in range(n_pos):
boxb = truth_boxes[j]
iou = box_iou(boxa, boxb)
if (iou > ovmax):
ovmax = iou
max_index = j
if (ovmax > 0.5):
if (gt_covered[max_index] == 0):
tp[i] = 1
gt_covered[max_index] = 1
if (len(boxes) != 0):
ap += sum(tp) * 1.0 / (len(boxes))
image_num += 1
rec += sum(gt_covered) * 1.0 / n_pos
return ap / image_num, rec / len(imageList)
def Acc(imageList,model,sample_number=5000,thresh=0.3):
classes = 20
side = 7
count = 0
ap = 0.0
rec = 0.0
image_num = 0
for image in imageList:
count += 1
if(count % 500 == 0):
print 'Image number:', count
#Get prediction from neural network
img = crop_detection(image.imgPath,new_width=448,new_height=448)
img = np.expand_dims(img, axis=0)
out = model.predict(img)
predictions = out[0]
#Post process predicting results
boxes = convert_yolo_detections(predictions)
boxes = do_nms_sort(boxes)
#pack ground truth boxes to true_boxes
truth_boxes = []
for i in range(side):
for j in range(side):
box_i_j = image.boxes[i][j]
if(box_i_j.has_obj):
for obj in box_i_j.objs:
new_box = box(classes)
new_box.row = i
new_box.col = j
new_box.x = (j+obj.x)*64
new_box.y = (i+obj.y)*64
new_box.h = (obj.h ** 2) * 448
new_box.w = (obj.w ** 2) * 448
new_box.class_num = obj.class_num
truth_boxes.append(new_box)
#assign detections to ground truth
n_pos = len(truth_boxes)
tp = np.zeros(len(boxes))
gt_covered = np.zeros(n_pos)
for i in range(len(boxes)):
#find the one ground truth with highest iou
boxa = boxes[i]
ovmax = -1
max_index = 0
for j in range(n_pos):
boxb = truth_boxes[j]
iou = box_iou(boxa,boxb)
if(iou > ovmax):
ovmax = iou
max_index = j
if(ovmax > 0.5):
if(gt_covered[max_index] == 0):
tp[i] = 1
gt_covered[max_index] = 1
if(len(boxes) != 0):
ap += sum(tp)*1.0/(len(boxes))
image_num += 1
rec += sum(gt_covered)*1.0/n_pos
return ap / image_num,rec / len(imageList)
