def stack_different_imgs(imgs, cnt):
stack = list()
for img in imgs:
img_tmp, bb = ia.get_fragment(img, cnt)
stack.append(img_tmp)
img_stacked = np.dstack(stack)
return img_stacked, bb
#read current frame
ret,img = cap.read()
#if current frame does not exist break
if not ret:
break
#make a first prediction for the whole current frame
img_whole = nh.evaluate_plus_resizing(net_whole,img,dim_whole,device=device,normalize=nh.pytorch_normalize_3)
#find contour of the shape of the first prediction
cnts = ia.filtered_contours(img_whole)
#make a second more accurate prediciton for each object of the first prediction
#and combine them into a new image "img_whole_plus_fragment"
img_whole_plus_fragment = np.zeros(img_whole.shape,dtype=np.uint8)
for cnt in cnts:
img_tmp, bb = ia.get_fragment(img,cnt)
pos = nh.evaluate_plus_resizing(net_fragment,img_tmp,dim_fragment,device=device,normalize=nh.pytorch_normalize_3)
neg = nh.evaluate_plus_resizing(net_fragment_negative,img_tmp,dim_fragment,device=device,normalize=nh.pytorch_normalize_3)
img_fragment = combine_pos_neg(pos,neg)
img_fragment = cv2.erode(img_fragment,ia.get_kernel(8))
img_fragment = ia.remove_filtered_contours(img_fragment)
img_whole_plus_fragment[bb[0]:bb[1],bb[2]:bb[3]] += img_fragment
#find all the object from the second prediction
cnts = ia.filtered_contours(img_whole_plus_fragment)
#if mice are seperated
if len(cnts) == 2:
#determine which mice is which
centers = ia.center_of_contours(cnts)
if get_distance(centers[0],centers_last[0]) + get_distance(centers[1],centers_last[1]) >= get_distance(centers[0],centers_last[1]) + get_distance(centers[1],centers_last[0]):
print('predicting behavior')
while 1:
ret, img = cap.read()
if not ret:
break
#dm.lucid_img(img,'img')
img_whole = nh.evaluate_plus_resizing(net_whole,
img,
dim_whole,
device=device,
normalize=nh.pytorch_normalize_3)
cnts = ia.filtered_contours(img_whole)
img_whole_plus_fragment = np.zeros(img_whole.shape, dtype=np.uint8)
for cnt in cnts:
img_tmp, bb = ia.get_fragment(img, cnt)
pos = nh.evaluate_plus_resizing(net_fragment,
img_tmp,
dim_fragment,
device=device,
normalize=nh.pytorch_normalize_3)
neg = nh.evaluate_plus_resizing(net_fragment_negative,
img_tmp,
dim_fragment,
device=device,
normalize=nh.pytorch_normalize_3)
img_fragment = combine_pos_neg(pos, neg)
img_fragment = ia.remove_filtered_contours(img_fragment)
img_whole_plus_fragment[bb[0]:bb[1], bb[2]:bb[3]] += img_fragment
img1 = cv2.bitwise_and(img1, img1, mask=bin1)
img1 = cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY)
img2 = cv2.bitwise_and(img2, img2, mask=bin2)
img2 = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY)
#remove part of the mouse in the back which is overlapping with the mice in the front and add both images together
img2 = cv2.bitwise_and(img2, img2, mask=cv2.bitwise_not(bin1))
img_c = img1 + img2
#apply histogramm equalization
img_c = cv2.equalizeHist(img_c)
#get the close up image of both overlapping mice
cnt1 = ia.get_max_contour(bin1)
cnt2 = ia.get_max_contour(bin2)
img_c1, bb1 = ia.get_fragment(img_c, cnt1)
img_c2, bb2 = ia.get_fragment(img_c, cnt2)
#create a combined binary image
bin_g1 = bin1
bin_g2 = bin2
bin_c = np.stack((bin_g1, bin_g2), axis=2)
#get the same close up image of the binary iamge
bin_c1, bb1 = ia.get_fragment(bin_c, cnt1)
bin_c2, bb2 = ia.get_fragment(bin_c, cnt2)
#reshape into consistent shape
img_c1 = cv2.resize(img_c1, dim, interpolation=cv2.INTER_AREA)
bin_c1 = cv2.resize(bin_c1, dim, interpolation=cv2.INTER_AREA)
img_c2 = cv2.resize(img_c2, dim, interpolation=cv2.INTER_AREA)