import numpy as np
#from builtins import range
boxtype = np.dtype([('x', np.float32), ('y', np.float32), ('w', np.float32), ('h', np.float32)])
def max_index(a, n): # float* and int
if n <= 0:
return -1
i = 0
max_i = 0
max_v = a[0]
for i in range(1, n):
if a[i] > max_v:
max_v = a[i]
max_i = i
return max_i
def resize_image(data, size):
"""
resizes given image (as an array in which channels are given in last dimension) to given size
"""
import numpy as np
import scipy
(w, h, channels) = data.shape
(target_w, target_h) = size
scale_w = 1.0*(w - 1)/(target_w - 1)
scale_h = 1.0*(h - 1)/(target_h - 1)
part = np.zeros((target_w, h, 3))
for c in range(channels):
arr = np.asarray([[ min(i*scale_w, w - 1) for i in range(target_w)]])
for j in range(h):
part[:, j, c] = scipy.ndimage.map_coordinates(data[:, j, c], arr, order=1)
resized = np.zeros((target_w, target_h, channels))
for c in range(channels):
arr = np.asarray([[min(j*scale_h, h - 1) for j in range(target_h)]])
for i in range(target_w):
resized[i, :, c] = scipy.ndimage.map_coordinates(part[i, :, c], arr, order=1)
return resized
def letterbox_image(img_data, size):
"""
exactly the same as letterbox_image from image.c from darknet sources:
resizes image (given as a RGB 0-1 float values) and fills empty space with 0.5
"""
import numpy as np
import scipy
(target_w, target_h) = size
new_w = target_w
new_h = target_h
(w, h, c) = img_data.shape
if 1.0*target_w/w < 1.0*target_h/h:
new_w = target_w
new_h = (h*target_w)//w
else:
new_h = target_h
new_w = (w*target_h)//h
resized = resize_image(img_data, (new_w, new_h))
assert resized.shape == (new_w, new_h, c)
x0, y0 = (target_w - new_w)//2, (target_h - new_h)//2 # upper left corner of resized image
resized = np.pad(
resized,
[
(max(x0 - 1, 0), max(target_w - new_w - max(x0 - 1, 0), 0)),
(max(y0 - 1, 0), max(target_h - new_h - max(y0 - 1, 0), 0)),
(0, 0)
], 'constant', constant_values=0.5)
return resized
def load_img(img_path, target_size):
"""
loads the image the same way darknet does, processes it and returns it (as array).
uses PIL, like keras.preprocessing.image module.
This loads image in RGB format.
"""
from PIL import Image as pil_image
import keras.backend as K
import numpy as np
img = pil_image.open(img_path)
# TODO: check format and convert to RGB
#resize
x = np.asarray(img, dtype=K.floatx())/255.0
#print(x[0,0,0], x[1,0,0], x[0,1,0], x[1,1,0], img.mode)
x = letterbox_image(x, target_size)
return x
def draw_box_width(img, left, top, right, bottom, r, g, b):
from PIL import ImageDraw
draw = ImageDraw.Draw(img)
draw.rectangle([left, top, right, bottom], outline=(r,g,b))
#draw.text((10, 25-10), "world")
def draw_label(img, left, top, right, bottom, r, g, b, label):
from PIL import ImageDraw
draw = ImageDraw.Draw(img)
print(repr(label))
(w, h) = draw.textsize(label.encode())
draw.rectangle([left, top-h, left+w, top], fill=(r,g,b))
draw.text((left, max(top-h, 0)), label.encode(), fill=(255,255,255))
def draw_detections(img, thresh, boxes, probs, names, alphabet, classes, w=13, h=13, n=5):
for k in range(n):
for j in range(h):
for i in range(w):
class0 = max_index(probs[i, j, k, :], classes)
prob = probs[i, j, k, class0]
if prob <= thresh:
continue
# prob > thresh
#width = h*0.012
#printf("%s: %.0f%%\n", names[class0], prob*100);
print("%s: %5i%%" % (names[class0], int(prob*100)))
offset = class0*123457 % classes
color = ((256*(class0+1)/(classes+1))+ (class0+1)/(classes+1))*256+(class0+1)/(classes+1)+offset
red = int(color/256/256)
blue = int(color %256)
green = int((color/256)%256)
#red = get_color(2,offset,classes);
#green = get_color(1,offset,classes);
#blue = get_color(0,offset,classes);
b = boxes[i, j, k]
left = int((b["x"] - b["w"]/2.0)*img.width)
right = int((b["x"] + b["w"]/2.0)*img.width)
top = int((b["y"] - b["h"]/2)*img.height)
bottom = int((b["y"] + b["h"]/2.0)*img.height)
left = max(left, 0)
right = min(right, img.width - 1)
top = max(top, 0)
bottom = min(bottom, img.height - 1)
draw_box_width(img, left, top, right, bottom, red, green, blue)
draw_label(img, left, top, right, bottom, red, green, blue, names[class0])
pass
