# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
from scipy.special import softmax
def hard_nms(box_scores, iou_threshold, top_k=-1, candidate_size=200):
"""
Args:
box_scores (N, 5): boxes in corner-form and probabilities.
iou_threshold: intersection over union threshold.
top_k: keep top_k results. If k <= 0, keep all the results.
candidate_size: only consider the candidates with the highest scores.
Returns:
picked: a list of indexes of the kept boxes
"""
scores = box_scores[:, -1]
boxes = box_scores[:, :-1]
picked = []
indexes = np.argsort(scores)
indexes = indexes[-candidate_size:]
while len(indexes) > 0:
current = indexes[-1]
picked.append(current)
if 0 < top_k == len(picked) or len(indexes) == 1:
break
current_box = boxes[current, :]
indexes = indexes[:-1]
rest_boxes = boxes[indexes, :]
iou = iou_of(
rest_boxes,
np.expand_dims(current_box, axis=0),
)
indexes = indexes[iou <= iou_threshold]
return box_scores[picked, :]
def iou_of(boxes0, boxes1, eps=1e-5):
"""Return intersection-over-union (Jaccard index) of boxes.
Args:
boxes0 (N, 4): ground truth boxes.
boxes1 (N or 1, 4): predicted boxes.
eps: a small number to avoid 0 as denominator.
Returns:
iou (N): IoU values.
"""
overlap_left_top = np.maximum(boxes0[..., :2], boxes1[..., :2])
overlap_right_bottom = np.minimum(boxes0[..., 2:], boxes1[..., 2:])
overlap_area = area_of(overlap_left_top, overlap_right_bottom)
area0 = area_of(boxes0[..., :2], boxes0[..., 2:])
area1 = area_of(boxes1[..., :2], boxes1[..., 2:])
return overlap_area / (area0 + area1 - overlap_area + eps)
def area_of(left_top, right_bottom):
"""Compute the areas of rectangles given two corners.
Args:
left_top (N, 2): left top corner.
right_bottom (N, 2): right bottom corner.
Returns:
area (N): return the area.
"""
hw = np.clip(right_bottom - left_top, 0.0, None)
return hw[..., 0] * hw[..., 1]
def calculate_containment(boxes0, boxes1):
"""
Calculate the containment of the boxes.
Args:
boxes0 (N, 4): ground truth boxes.
boxes1 (N or 1, 4): predicted boxes.
Returns:
containment (N): containment values.
"""
overlap_left_top = np.maximum(boxes0[..., :2], boxes1[..., :2])
overlap_right_bottom = np.minimum(boxes0[..., 2:], boxes1[..., 2:])
overlap_area = area_of(overlap_left_top, overlap_right_bottom)
area0 = area_of(boxes0[..., :2], boxes0[..., 2:])
area1 = area_of(boxes1[..., :2], boxes1[..., 2:])
return overlap_area / np.minimum(area0, np.expand_dims(area1, axis=0))
class PicoDetPostProcess(object):
"""
Args:
input_shape (int): network input image size
ori_shape (int): ori image shape of before padding
scale_factor (float): scale factor of ori image
enable_mkldnn (bool): whether to open MKLDNN
"""
def __init__(
self,
layout_dict_path,
strides=[8, 16, 32, 64],
score_threshold=0.4,
nms_threshold=0.5,
nms_top_k=1000,
keep_top_k=100,
):
self.labels = self.load_layout_dict(layout_dict_path)
self.strides = strides
self.score_threshold = score_threshold
self.nms_threshold = nms_threshold
self.nms_top_k = nms_top_k
self.keep_top_k = keep_top_k
def load_layout_dict(self, layout_dict_path):
with open(layout_dict_path, "r", encoding="utf-8") as fp:
labels = fp.readlines()
return [label.strip("\n") for label in labels]
def warp_boxes(self, boxes, ori_shape):
"""Apply transform to boxes"""
width, height = ori_shape[1], ori_shape[0]
n = len(boxes)
if n:
# warp points
xy = np.ones((n * 4, 3))
xy[:, :2] = boxes[:, [0, 1, 2, 3, 0, 3, 2, 1]].reshape(
n * 4, 2
) # x1y1, x2y2, x1y2, x2y1
# xy = xy @ M.T # transform
xy = (xy[:, :2] / xy[:, 2:3]).reshape(n, 8) # rescale
# create new boxes
x = xy[:, [0, 2, 4, 6]]
y = xy[:, [1, 3, 5, 7]]
xy = (
np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
)
# clip boxes
xy[:, [0, 2]] = xy[:, [0, 2]].clip(0, width)
xy[:, [1, 3]] = xy[:, [1, 3]].clip(0, height)
return xy.astype(np.float32)
else:
return boxes
def img_info(self, ori_img, img):
origin_shape = ori_img.shape
resize_shape = img.shape
im_scale_y = resize_shape[2] / float(origin_shape[0])
im_scale_x = resize_shape[3] / float(origin_shape[1])
scale_factor = np.array([im_scale_y, im_scale_x], dtype=np.float32)
img_shape = np.array(img.shape[2:], dtype=np.float32)
input_shape = np.array(img).astype("float32").shape[2:]
ori_shape = np.array((img_shape,)).astype("float32")
scale_factor = np.array((scale_factor,)).astype("float32")
return ori_shape, input_shape, scale_factor
def __call__(self, ori_img, img, preds):
scores, raw_boxes = preds["boxes"], preds["boxes_num"]
batch_size = raw_boxes[0].shape[0]
reg_max = int(raw_boxes[0].shape[-1] / 4 - 1)
out_boxes_num = []
out_boxes_list = []
results = []
ori_shape, input_shape, scale_factor = self.img_info(ori_img, img)
for batch_id in range(batch_size):
# generate centers
decode_boxes = []
select_scores = []
for stride, box_distribute, score in zip(self.strides, raw_boxes, scores):
box_distribute = box_distribute[batch_id]
score = score[batch_id]
# centers
fm_h = input_shape[0] / stride
fm_w = input_shape[1] / stride
h_range = np.arange(fm_h)
w_range = np.arange(fm_w)
ww, hh = np.meshgrid(w_range, h_range)
ct_row = (hh.flatten() + 0.5) * stride
ct_col = (ww.flatten() + 0.5) * stride
center = np.stack((ct_col, ct_row, ct_col, ct_row), axis=1)
# box distribution to distance
reg_range = np.arange(reg_max + 1)
box_distance = box_distribute.reshape((-1, reg_max + 1))
box_distance = softmax(box_distance, axis=1)
box_distance = box_distance * np.expand_dims(reg_range, axis=0)
box_distance = np.sum(box_distance, axis=1).reshape((-1, 4))
box_distance = box_distance * stride
# top K candidate
topk_idx = np.argsort(score.max(axis=1))[::-1]
topk_idx = topk_idx[: self.nms_top_k]
center = center[topk_idx]
score = score[topk_idx]
box_distance = box_distance[topk_idx]
# decode box
decode_box = center + [-1, -1, 1, 1] * box_distance
select_scores.append(score)
decode_boxes.append(decode_box)
# nms
bboxes = np.concatenate(decode_boxes, axis=0)
confidences = np.concatenate(select_scores, axis=0)
picked_box_probs = []
picked_labels = []
for class_index in range(0, confidences.shape[1]):
probs = confidences[:, class_index]
mask = probs > self.score_threshold
probs = probs[mask]
if probs.shape[0] == 0:
continue
subset_boxes = bboxes[mask, :]
box_probs = np.concatenate([subset_boxes, probs.reshape(-1, 1)], axis=1)
box_probs = hard_nms(
box_probs,
iou_threshold=self.nms_threshold,
top_k=self.keep_top_k,
)
picked_box_probs.append(box_probs)
picked_labels.extend([class_index] * box_probs.shape[0])
if len(picked_box_probs) == 0:
out_boxes_list.append(np.empty((0, 4)))
out_boxes_num.append(0)
else:
picked_box_probs = np.concatenate(picked_box_probs)
# resize json boxes
picked_box_probs[:, :4] = self.warp_boxes(
picked_box_probs[:, :4], ori_shape[batch_id]
)
im_scale = np.concatenate(
[scale_factor[batch_id][::-1], scale_factor[batch_id][::-1]]
)
picked_box_probs[:, :4] /= im_scale
# clas score box
out_boxes_list.append(
np.concatenate(
[
np.expand_dims(np.array(picked_labels), axis=-1),
np.expand_dims(picked_box_probs[:, 4], axis=-1),
picked_box_probs[:, :4],
],
axis=1,
)
)
out_boxes_num.append(len(picked_labels))
out_boxes_list = np.concatenate(out_boxes_list, axis=0)
out_boxes_num = np.asarray(out_boxes_num).astype(np.int32)
for dt in out_boxes_list:
clsid, bbox, score = int(dt[0]), dt[2:], dt[1]
label = self.labels[clsid]
result = {"bbox": bbox, "label": label, "score": score}
results.append(result)
# Handle conflict where a box is simultaneously recognized as multiple labels.
# Use IoU to find similar boxes. Prioritize labels as table, text, and others when deduplicate similar boxes.
bboxes = np.array([x["bbox"] for x in results])
duplicate_idx = list()
for i in range(len(results)):
if i in duplicate_idx:
continue
containments = calculate_containment(bboxes, bboxes[i, ...])
overlaps = np.where(containments > 0.5)[0]
if len(overlaps) > 1:
table_box = [x for x in overlaps if results[x]["label"] == "table"]
if len(table_box) > 0:
keep = sorted(
[(x, results[x]) for x in table_box],
key=lambda x: x[1]["score"],
reverse=True,
)[0][0]
else:
keep = sorted(
[(x, results[x]) for x in overlaps],
key=lambda x: x[1]["score"],
reverse=True,
)[0][0]
duplicate_idx.extend([x for x in overlaps if x != keep])
results = [x for i, x in enumerate(results) if i not in duplicate_idx]
return results