def extract_region(field, model, x, y, l, box_length, stride, threshold=0.97, prune=False):
im = rgb2grey(field[x:x + l, y:y + l]) # transform from color to grey
im = resize(im, (im.shape[0] * 3, im.shape[1] * 3)) ##scale it up, hideous i know, but you can't train on this 8x8 pixels. plus the ground_truth was scaled....
##collect all the boxes and the probabilities.
box, prob = sliding_window_count_vectorised(im, model, length=box_length, stride=stride,
probability_threshold=threshold)
# no probs, no boxes, bail!
if len(prob) is 0:
return box, prob
# test code to view boxes etc.
#to_draw = draw_boxes(grey2rgb(im), box)
#plt.imshow(to_draw)
#plt.show()
#plt.show()
#imsave("./hand_made/pos_4_truth.png", to_draw)
if prune:
box, prob = non_max_suppression_fast(box, prob, 0.18)
# need to shift the boxes co-ords, and downsample them relative to the global region in the field x,y
box = box.astype(float)
box /= 3.0
box += np.array([x, y, x, y])
return box, prob
def evaluate_whole_field(output_dir, field, model, l=250, stride=5, prune=True):
#run through the image cutting off 1k squres.
box_length = 20
h, w = field.shape[:2]
##load the main three variables.
start = np.array([0,0])
if os.path.exists(output_dir+"loop_vars.npy"):
start = np.load(output_dir+"loop_vars.npy")
boxes = None
if os.path.exists(output_dir+"boxes.npy"):
boxes = np.load(output_dir+"boxes.npy")
else:
boxes = np.zeros((1, 4))
probs = None
if os.path.exists(output_dir+"probs.npy"):
probs = np.load(output_dir+"probs.npy")
else:
probs = np.zeros((1))
#we take off box length in case of an overlap.
for x in range(start[0], h, l-box_length):
for y in range(start[1], w, l-box_length):
print("%d, %d" % (x,y))
# Prevent doing all this work for all black squares
if np.max(field[x:x+l,y:y+l]) == 0:
continue
np.save(output_dir+"loop_vars.npy", np.array([x, y]))
box, prob = extract_region(field, model, x, y, l, box_length, stride, threshold=0.90, prune=prune)
if len(box) is not 0:
boxes = np.vstack((boxes,box))
probs = np.hstack((probs,prob))
#save the values for loading.
np.save(output_dir+"boxes.npy", boxes)
np.save(output_dir+"probs.npy", probs)
start = np.array([x, 0])
np.save(output_dir+"loop_vars.npy", start)
#set the loop vars to done.
np.save(output_dir + "loop_vars.npy", np.array([h, w]))
##prune the overlapping boxes.
if not prune:
boxes, probs = non_max_suppression_fast(boxes, probs, 0.18)
np.save(output_dir + "pruned_boxes.npy", boxes)
np.save(output_dir + "pruned_probs.npy", probs)
print(boxes.shape)
imsave(name+"_lettuce_count_" + str(boxes.shape[0]) + ".png", draw_boxes(grey2rgb(field), boxes, color=(255,0,0)))
def prune_boxes(name,overlap_coefficient=0.18):
boxes = np.load("boxes.npy")
print(boxes.shape)
probs = np.load("probs.npy")
boxes, probs = non_max_suppression_fast(boxes, probs, 0.18)
np.save(name + "/pruned_boxes.npy", boxes)
np.save(name + "/pruned_probs.npy", probs)
print(boxes.shape)
boxes = np.save(name + "/pruned_boxes.npy", boxes)
imsave(name+"_lettuce_count_" + str(boxes.shape[0]) + ".png", draw_boxes(grey2rgb(whole_field), boxes, color=(255,0,0)))
def evaluate_whole_field(output_dir,
field,
model,
l=250,
stride=5,
prune=True):
#run through the image cutting off 1k squres.
box_length = 20
h, w = field.shape[:2]
##load the main three variables.
start = np.array([0, 0])
if os.path.exists(output_dir + "loop_vars.npy"):
start = np.load(output_dir + "loop_vars.npy")
boxes = None
if os.path.exists(output_dir + "boxes.npy"):
boxes = np.load(output_dir + "boxes.npy")
else:
boxes = np.zeros((1, 4))
probs = None
if os.path.exists(output_dir + "probs.npy"):
probs = np.load(output_dir + "probs.npy")
else:
probs = np.zeros((1))
#we take off box length in case of an overlap.
for x in range(start[0], h, l - box_length):
for y in range(start[1], w, l - box_length):
print("%d, %d" % (x, y))
np.save(output_dir + "loop_vars.npy", np.array([x, y]))
box, prob = extract_region(field,
model,
x,
y,
l,
box_length,
stride,
threshold=0.90,
prune=prune)
if len(box) is not 0:
boxes = np.vstack((boxes, box))
probs = np.hstack((probs, prob))
#save the values for loading.
np.save(output_dir + "boxes.npy", boxes)
np.save(output_dir + "probs.npy", probs)
start = np.array([x, 0])
np.save(output_dir + "loop_vars.npy", start)
#set the loop vars to done.
np.save(output_dir + "loop_vars.npy", np.array([h, w]))
##prune the overlapping boxes.
if not prune:
boxes, probs = non_max_suppression_fast(boxes, probs, 0.18)
np.save(output_dir + "pruned_boxes.npy", boxes)
np.save(output_dir + "pruned_probs.npy", probs)
print(boxes.shape)
