Hi!
First, thanks for the implementation, nice work!
I've that in get_image_boxes PIL.Image is casted to np.array multiple times for every bounding box found. AFAIK, the fix doesn't break anything, results look the same.
Just for the perspective, I've tested on a large image (around 20MB in size):
Timer unit: 1 s
Total time: 13.8626 s
Function: get_image_boxes at line 127
Line # Hits Time Per Hit % Time Line Contents
==============================================================
127 def get_image_boxes(bounding_boxes, img, size=24):
128 """Cut out boxes from the image.
129
130 Arguments:
131 bounding_boxes: a float numpy array of shape [n, 5].
132 img: an instance of PIL.Image.
133 size: an integer, size of cutouts.
134
135 Returns:
136 a float numpy array of shape [n, 3, size, size].
137 """
138 2 0.0 0.0 0.0 num_boxes = len(bounding_boxes)
139 2 0.0 0.0 0.0 width, height = img.size
140
141 2 0.0 0.0 0.0 [dy, edy, dx, edx, y, ey, x, ex, w, h] = correct_bboxes(bounding_boxes, width, height)
142 2 0.0 0.0 0.0 img_boxes = np.zeros((num_boxes, 3, size, size), 'float32')
143
144 397 0.0 0.0 0.0 for i in range(num_boxes):
145 395 13.7 0.0 99.0 img_array = np.asarray(img, 'uint8')
146 395 0.0 0.0 0.1 img_box = np.zeros((h[i], w[i], 3), 'uint8')
147 img_box[dy[i]:(edy[i] + 1), dx[i]:(edx[i] + 1), :] =\
148 395 0.0 0.0 0.2 img_array[y[i]:(ey[i] + 1), x[i]:(ex[i] + 1), :]
149
150 # resize
151 395 0.0 0.0 0.3 img_box = Image.fromarray(img_box)
152 395 0.0 0.0 0.1 img_box = img_box.resize((size, size), Image.BILINEAR)
153 395 0.0 0.0 0.1 img_box = np.asarray(img_box, 'float32')
154
155 395 0.0 0.0 0.1 img_boxes[i, :, :, :] = _preprocess(img_box)
156
157 2 0.0 0.0 0.0 return img_boxes
Hi! First, thanks for the implementation, nice work! I've that in
get_image_boxesPIL.Image is casted tonp.arraymultiple times for every bounding box found. AFAIK, the fix doesn't break anything, results look the same. Just for the perspective, I've tested on a large image (around 20MB in size):With the fix, total time is reduces to ~1 second.