deploy: fb8481effdf5a0b23ff86fad414906046d7620bd
Browse files- layout-unreadability.py +75 -15
layout-unreadability.py
CHANGED
|
@@ -15,7 +15,32 @@ Computes the non-flatness of regions that text elements are solely put on, refer
|
|
| 15 |
"""
|
| 16 |
|
| 17 |
_KWARGS_DESCRIPTION = """\
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
"""
|
| 20 |
|
| 21 |
_CITATION = """\
|
|
@@ -35,8 +60,8 @@ ReqType = Literal["pil2cv", "cv2pil"]
|
|
| 35 |
class LayoutUnreadability(evaluate.Metric):
|
| 36 |
def __init__(
|
| 37 |
self,
|
| 38 |
-
canvas_width: int,
|
| 39 |
-
canvas_height: int,
|
| 40 |
text_label_index: int = 1,
|
| 41 |
decoration_label_index: int = 3,
|
| 42 |
**kwargs,
|
|
@@ -98,6 +123,8 @@ class LayoutUnreadability(evaluate.Metric):
|
|
| 98 |
def load_image_canvas(
|
| 99 |
self,
|
| 100 |
filepath: Union[os.PathLike, List[os.PathLike]],
|
|
|
|
|
|
|
| 101 |
) -> npt.NDArray[np.float64]:
|
| 102 |
if isinstance(filepath, list):
|
| 103 |
assert len(filepath) == 1, filepath
|
|
@@ -105,8 +132,8 @@ class LayoutUnreadability(evaluate.Metric):
|
|
| 105 |
|
| 106 |
canvas_pil = Image.open(filepath) # type: ignore
|
| 107 |
canvas_pil = canvas_pil.convert("RGB") # type: ignore
|
| 108 |
-
if canvas_pil.size != (
|
| 109 |
-
canvas_pil = canvas_pil.resize((
|
| 110 |
|
| 111 |
canvas_pil = self.img_to_g_xy(canvas_pil)
|
| 112 |
assert isinstance(canvas_pil, PilImage)
|
|
@@ -115,20 +142,24 @@ class LayoutUnreadability(evaluate.Metric):
|
|
| 115 |
return canvas_arr
|
| 116 |
|
| 117 |
def get_rid_of_invalid(
|
| 118 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
) -> npt.NDArray[np.int64]:
|
| 120 |
assert len(predictions) == len(gold_labels)
|
| 121 |
|
| 122 |
-
w =
|
| 123 |
-
h =
|
| 124 |
|
| 125 |
for i, prediction in enumerate(predictions):
|
| 126 |
for j, b in enumerate(prediction):
|
| 127 |
xl, yl, xr, yr = b
|
| 128 |
xl = max(0, xl)
|
| 129 |
yl = max(0, yl)
|
| 130 |
-
xr = min(
|
| 131 |
-
yr = min(
|
| 132 |
if abs((xr - xl) * (yr - yl)) < w * h * 10:
|
| 133 |
if gold_labels[i, j]:
|
| 134 |
gold_labels[i, j] = 0
|
|
@@ -140,15 +171,42 @@ class LayoutUnreadability(evaluate.Metric):
|
|
| 140 |
predictions: Union[npt.NDArray[np.float64], List[List[float]]],
|
| 141 |
gold_labels: Union[npt.NDArray[np.int64], List[int]],
|
| 142 |
image_canvases: List[os.PathLike],
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
predictions = np.array(predictions)
|
| 145 |
gold_labels = np.array(gold_labels)
|
| 146 |
|
| 147 |
-
predictions[:, :, ::2] *=
|
| 148 |
-
predictions[:, :, 1::2] *=
|
| 149 |
|
| 150 |
gold_labels = self.get_rid_of_invalid(
|
| 151 |
-
predictions=predictions,
|
|
|
|
|
|
|
|
|
|
| 152 |
)
|
| 153 |
score = 0.0
|
| 154 |
|
|
@@ -159,16 +217,18 @@ class LayoutUnreadability(evaluate.Metric):
|
|
| 159 |
for prediction, gold_label, image_canvas in it:
|
| 160 |
canvas_arr = self.load_image_canvas(
|
| 161 |
image_canvas,
|
|
|
|
|
|
|
| 162 |
)
|
| 163 |
cal_mask = np.zeros_like(canvas_arr)
|
| 164 |
|
| 165 |
prediction = np.array(prediction, dtype=int)
|
| 166 |
gold_label = np.array(gold_label, dtype=int)
|
| 167 |
|
| 168 |
-
is_text = (gold_label ==
|
| 169 |
prediction_text = prediction[is_text]
|
| 170 |
|
| 171 |
-
is_decoration = (gold_label ==
|
| 172 |
prediction_deco = prediction[is_decoration]
|
| 173 |
|
| 174 |
for mp in prediction_text:
|
|
|
|
| 15 |
"""
|
| 16 |
|
| 17 |
_KWARGS_DESCRIPTION = """\
|
| 18 |
+
Args:
|
| 19 |
+
predictions (`list` of `list` of `float`): A list of lists of floats representing normalized `ltrb`-format bounding boxes.
|
| 20 |
+
gold_labels (`list` of `list` of `int`): A list of lists of integers representing class labels.
|
| 21 |
+
image_canvases (`list` of `str`): A list of file paths to canvas images (background images).
|
| 22 |
+
canvas_width (`int`, *optional*): Width of the canvas in pixels. Can be provided at initialization or during computation.
|
| 23 |
+
canvas_height (`int`, *optional*): Height of the canvas in pixels. Can be provided at initialization or during computation.
|
| 24 |
+
text_label_index (`int`, *optional*, defaults to 1): The label index for text elements.
|
| 25 |
+
decoration_label_index (`int`, *optional*, defaults to 3): The label index for decoration (underlay) elements.
|
| 26 |
+
|
| 27 |
+
Returns:
|
| 28 |
+
float: The unreadability score measuring the non-flatness of regions where text elements are placed. Computed using gradient analysis (Sobel operator) on the canvas image. Lower values indicate better readability (text on flatter/cleaner backgrounds).
|
| 29 |
+
|
| 30 |
+
Examples:
|
| 31 |
+
>>> import evaluate
|
| 32 |
+
>>> metric = evaluate.load("creative-graphic-design/layout-unreadability")
|
| 33 |
+
>>> predictions = [[[0.1, 0.1, 0.5, 0.3], [0.6, 0.6, 0.9, 0.8]]]
|
| 34 |
+
>>> gold_labels = [[1, 2]] # 1 is text, 2 is other element
|
| 35 |
+
>>> image_canvases = ["/path/to/canvas.png"]
|
| 36 |
+
>>> result = metric.compute(
|
| 37 |
+
... predictions=predictions,
|
| 38 |
+
... gold_labels=gold_labels,
|
| 39 |
+
... image_canvases=image_canvases,
|
| 40 |
+
... canvas_width=512,
|
| 41 |
+
... canvas_height=512
|
| 42 |
+
... )
|
| 43 |
+
>>> print(f"Unreadability score: {result:.4f}")
|
| 44 |
"""
|
| 45 |
|
| 46 |
_CITATION = """\
|
|
|
|
| 60 |
class LayoutUnreadability(evaluate.Metric):
|
| 61 |
def __init__(
|
| 62 |
self,
|
| 63 |
+
canvas_width: int | None = None,
|
| 64 |
+
canvas_height: int | None = None,
|
| 65 |
text_label_index: int = 1,
|
| 66 |
decoration_label_index: int = 3,
|
| 67 |
**kwargs,
|
|
|
|
| 123 |
def load_image_canvas(
|
| 124 |
self,
|
| 125 |
filepath: Union[os.PathLike, List[os.PathLike]],
|
| 126 |
+
canvas_width: int,
|
| 127 |
+
canvas_height: int,
|
| 128 |
) -> npt.NDArray[np.float64]:
|
| 129 |
if isinstance(filepath, list):
|
| 130 |
assert len(filepath) == 1, filepath
|
|
|
|
| 132 |
|
| 133 |
canvas_pil = Image.open(filepath) # type: ignore
|
| 134 |
canvas_pil = canvas_pil.convert("RGB") # type: ignore
|
| 135 |
+
if canvas_pil.size != (canvas_width, canvas_height):
|
| 136 |
+
canvas_pil = canvas_pil.resize((canvas_width, canvas_height)) # type: ignore
|
| 137 |
|
| 138 |
canvas_pil = self.img_to_g_xy(canvas_pil)
|
| 139 |
assert isinstance(canvas_pil, PilImage)
|
|
|
|
| 142 |
return canvas_arr
|
| 143 |
|
| 144 |
def get_rid_of_invalid(
|
| 145 |
+
self,
|
| 146 |
+
predictions: npt.NDArray[np.float64],
|
| 147 |
+
gold_labels: npt.NDArray[np.int64],
|
| 148 |
+
canvas_width: int,
|
| 149 |
+
canvas_height: int,
|
| 150 |
) -> npt.NDArray[np.int64]:
|
| 151 |
assert len(predictions) == len(gold_labels)
|
| 152 |
|
| 153 |
+
w = canvas_width / 100
|
| 154 |
+
h = canvas_height / 100
|
| 155 |
|
| 156 |
for i, prediction in enumerate(predictions):
|
| 157 |
for j, b in enumerate(prediction):
|
| 158 |
xl, yl, xr, yr = b
|
| 159 |
xl = max(0, xl)
|
| 160 |
yl = max(0, yl)
|
| 161 |
+
xr = min(canvas_width, xr)
|
| 162 |
+
yr = min(canvas_height, yr)
|
| 163 |
if abs((xr - xl) * (yr - yl)) < w * h * 10:
|
| 164 |
if gold_labels[i, j]:
|
| 165 |
gold_labels[i, j] = 0
|
|
|
|
| 171 |
predictions: Union[npt.NDArray[np.float64], List[List[float]]],
|
| 172 |
gold_labels: Union[npt.NDArray[np.int64], List[int]],
|
| 173 |
image_canvases: List[os.PathLike],
|
| 174 |
+
canvas_width: int | None = None,
|
| 175 |
+
canvas_height: int | None = None,
|
| 176 |
+
text_label_index: int | None = None,
|
| 177 |
+
decoration_label_index: int | None = None,
|
| 178 |
):
|
| 179 |
+
# パラメータの優先順位処理
|
| 180 |
+
canvas_width = canvas_width if canvas_width is not None else self.canvas_width
|
| 181 |
+
canvas_height = (
|
| 182 |
+
canvas_height if canvas_height is not None else self.canvas_height
|
| 183 |
+
)
|
| 184 |
+
text_label_index = (
|
| 185 |
+
text_label_index if text_label_index is not None else self.text_label_index
|
| 186 |
+
)
|
| 187 |
+
decoration_label_index = (
|
| 188 |
+
decoration_label_index
|
| 189 |
+
if decoration_label_index is not None
|
| 190 |
+
else self.decoration_label_index
|
| 191 |
+
)
|
| 192 |
+
|
| 193 |
+
if canvas_width is None or canvas_height is None:
|
| 194 |
+
raise ValueError(
|
| 195 |
+
"canvas_width and canvas_height must be provided either "
|
| 196 |
+
"at initialization or during computation"
|
| 197 |
+
)
|
| 198 |
+
|
| 199 |
predictions = np.array(predictions)
|
| 200 |
gold_labels = np.array(gold_labels)
|
| 201 |
|
| 202 |
+
predictions[:, :, ::2] *= canvas_width
|
| 203 |
+
predictions[:, :, 1::2] *= canvas_height
|
| 204 |
|
| 205 |
gold_labels = self.get_rid_of_invalid(
|
| 206 |
+
predictions=predictions,
|
| 207 |
+
gold_labels=gold_labels,
|
| 208 |
+
canvas_width=canvas_width,
|
| 209 |
+
canvas_height=canvas_height,
|
| 210 |
)
|
| 211 |
score = 0.0
|
| 212 |
|
|
|
|
| 217 |
for prediction, gold_label, image_canvas in it:
|
| 218 |
canvas_arr = self.load_image_canvas(
|
| 219 |
image_canvas,
|
| 220 |
+
canvas_width,
|
| 221 |
+
canvas_height,
|
| 222 |
)
|
| 223 |
cal_mask = np.zeros_like(canvas_arr)
|
| 224 |
|
| 225 |
prediction = np.array(prediction, dtype=int)
|
| 226 |
gold_label = np.array(gold_label, dtype=int)
|
| 227 |
|
| 228 |
+
is_text = (gold_label == text_label_index).reshape(-1)
|
| 229 |
prediction_text = prediction[is_text]
|
| 230 |
|
| 231 |
+
is_decoration = (gold_label == decoration_label_index).reshape(-1)
|
| 232 |
prediction_deco = prediction[is_decoration]
|
| 233 |
|
| 234 |
for mp in prediction_text:
|