File size: 7,173 Bytes
5f631c6 ab35335 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 |
# SPDX-FileCopyrightText: Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
import numpy as np
import pandas as pd
import numpy.typing as npt
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
from typing import Dict, List, Tuple, Optional, Union
COLORS = [
"#003EFF",
"#FF8F00",
"#079700",
"#A123FF",
"#87CEEB",
"#FF5733",
"#C70039",
"#900C3F",
"#581845",
"#11998E",
]
def reformat_for_plotting(
boxes: npt.NDArray[np.float64],
labels: npt.NDArray[np.int_],
scores: npt.NDArray[np.float64],
shape: Tuple[int, int],
num_classes: int,
) -> Tuple[List[npt.NDArray[np.int_]], List[npt.NDArray[np.float64]]]:
"""
Reformat YOLOX predictions for plotting.
- Unnormalizes boxes to original image size.
- Reformats boxes to [xmin, ymin, width, height].
- Converts to list of boxes and scores per class.
Args:
boxes (np.ndarray [N, 4]): Array of bounding boxes in format [xmin, ymin, xmax, ymax].
labels (np.ndarray [N]): Array of labels.
scores (np.ndarray [N]): Array of confidence scores.
shape (tuple [2]): Shape of the image (height, width).
num_classes (int): Number of classes.
Returns:
list[np.ndarray[N]]: List of box bounding boxes per class.
list[np.ndarray[N]]: List of confidence scores per class.
"""
boxes_plot = boxes.copy()
boxes_plot[:, [0, 2]] *= shape[1]
boxes_plot[:, [1, 3]] *= shape[0]
boxes_plot = boxes_plot.astype(int)
boxes_plot[:, 2] -= boxes_plot[:, 0]
boxes_plot[:, 3] -= boxes_plot[:, 1]
boxes_plot = [boxes_plot[labels == c] for c in range(num_classes)]
confs = [scores[labels == c] for c in range(num_classes)]
return boxes_plot, confs
def plot_sample(
img: npt.NDArray[np.uint8],
boxes_list: List[npt.NDArray[np.int_]],
confs_list: List[npt.NDArray[np.float64]],
labels: List[str],
show_text: bool = True,
) -> None:
"""
Plots an image with bounding boxes.
Coordinates are expected in format [x_min, y_min, width, height].
Args:
img (numpy.ndarray): The input image to be plotted.
boxes_list (list[np.ndarray]): List of box bounding boxes per class.
confs_list (list[np.ndarray]): List of confidence scores per class.
labels (list): List of class labels.
show_text (bool, optional): Whether to show the text. Defaults to True.
"""
plt.imshow(img, cmap="gray")
plt.axis(False)
for boxes, confs, col, l in zip(boxes_list, confs_list, COLORS, labels):
for box_idx, box in enumerate(boxes):
# Better display around boundaries
h, w, _ = img.shape
box = np.copy(box)
box[:2] = np.clip(box[:2], 2, max(h, w))
box[2] = min(box[2], w - 2 - box[0])
box[3] = min(box[3], h - 2 - box[1])
rect = Rectangle(
(box[0], box[1]),
box[2],
box[3],
linewidth=2,
facecolor="none",
edgecolor=col,
)
plt.gca().add_patch(rect)
# Add class and index label with proper alignment
if show_text:
plt.text(
box[0], box[1],
f"{l}_{box_idx} conf={confs[box_idx]:.3f}",
color='white',
fontsize=8,
bbox=dict(facecolor=col, alpha=1, edgecolor=col, pad=0, linewidth=2),
verticalalignment='bottom',
horizontalalignment='left'
)
def reorder_boxes(
boxes: npt.NDArray[np.float64],
labels: npt.NDArray[np.int_],
classes: Optional[List[str]] = None,
scores: Optional[npt.NDArray[np.float64]] = None,
) -> Union[
Tuple[npt.NDArray[np.float64], npt.NDArray[np.int_]],
Tuple[npt.NDArray[np.float64], npt.NDArray[np.int_], npt.NDArray[np.float64]],
]:
"""
Reorder boxes, labels and scores by box coordinates.
Columns are sorted by x first, rows and cells are sorted by y first.
Args:
boxes (np.ndarray [N, 4]): Array of bounding boxes in format [xmin, ymin, xmax, ymax].
labels (np.ndarray [N]): Array of labels.
classes (list, optional): List of class labels. Defaults to None.
scores (np.ndarray [N], optional): Array of confidence scores. Defaults to None.
Returns:
np.ndarray [N, 4]: Ordered boxes in format [xmin, ymin, xmax, ymax].
np.ndarray [N]: Ordered labels.
np.ndarray [N]: Ordered scores if scores is not None.
"""
n_classes = labels.max() if classes is None else len(classes)
classes = labels.unique() if classes is None else classes
ordered_boxes, ordered_labels, ordered_scores = [], [], []
for c in range(n_classes):
boxes_class = boxes[labels == c]
if len(boxes_class):
# Reorder
sort = ["x0", "y0"] if classes[c] == "column" else ["y0", "x0"]
df_coords = pd.DataFrame({
"y0": np.round(boxes_class[:, 1] - boxes_class[:, 1].min(), 2),
"x0": np.round(boxes_class[:, 0] - boxes_class[:, 0].min(), 2),
})
idxs = df_coords.sort_values(sort).index
ordered_boxes.append(boxes_class[idxs])
ordered_labels.append(labels[labels == c][idxs])
if scores is not None:
ordered_scores.append(scores[labels == c][idxs])
ordered_boxes = np.concatenate(ordered_boxes)
ordered_labels = np.concatenate(ordered_labels)
if scores is not None:
ordered_scores = np.concatenate(ordered_scores)
return ordered_boxes, ordered_labels, ordered_scores
return ordered_boxes, ordered_labels
def postprocess_preds_table_structure(
preds: Dict[str, npt.NDArray],
threshold: float = 0.1,
class_labels: Optional[List[str]] = None,
reorder: bool = True,
) -> Tuple[npt.NDArray[np.float64], npt.NDArray[np.int_], npt.NDArray[np.float64]]:
"""
Post process predictions for table structure task.
- Applies thresholding
- Reorders boxes using the reading order
Args:
preds (dict): Predictions. Keys are "scores", "boxes", "labels".
threshold (float, optional): Threshold for the confidence scores. Defaults to 0.1.
class_labels (list, optional): List of class labels. Defaults to None.
reorder (bool, optional): Whether to apply reordering. Defaults to True.
Returns:
numpy.ndarray [N x 4]: Array of bounding boxes.
numpy.ndarray [N]: Array of labels.
numpy.ndarray [N]: Array of scores.
"""
boxes = preds["boxes"].cpu().numpy()
labels = preds["labels"].cpu().numpy()
scores = preds["scores"].cpu().numpy()
# Threshold
boxes = boxes[scores > threshold]
labels = labels[scores > threshold]
scores = scores[scores > threshold]
if len(boxes) > 0 and reorder:
boxes, labels, scores = reorder_boxes(boxes, labels, class_labels, scores)
return boxes, labels, scores
|