# 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 time
import os
import ast
import argparse
from typing import List, Tuple
import numpy as np
class LayoutBox(object):
def __init__(self, clsid: int, pos: List[float], confidence: float):
self.clsid = clsid
self.pos = pos
self.confidence = confidence
class PageDetectionResult(object):
def __init__(self, boxes: List[LayoutBox], image_path: str):
self.boxes = boxes
self.image_path = image_path
def argsparser():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--image_dir",
type=str,
default=None,
help="Dir of image file, `image_file` has a higher priority.")
parser.add_argument(
"--batch_size", type=int, default=1, help="batch_size for inference.")
parser.add_argument(
"--video_file",
type=str,
default=None,
help="Path of video file, `video_file` or `camera_id` has a highest priority."
)
parser.add_argument(
"--camera_id",
type=int,
default=-1,
help="device id of camera to predict.")
parser.add_argument(
"--threshold", type=float, default=0.5, help="Threshold of score.")
parser.add_argument(
"--output_dir",
type=str,
default="output",
help="Directory of output visualization files.")
parser.add_argument(
"--run_mode",
type=str,
default='paddle',
help="mode of running(paddle/trt_fp32/trt_fp16/trt_int8)")
parser.add_argument(
"--device",
type=str,
default='cpu',
help="Choose the device you want to run, it can be: CPU/GPU/XPU/NPU, default is CPU."
)
parser.add_argument(
"--use_gpu",
type=ast.literal_eval,
default=False,
help="Deprecated, please use `--device`.")
parser.add_argument(
"--run_benchmark",
type=ast.literal_eval,
default=False,
help="Whether to predict a image_file repeatedly for benchmark")
parser.add_argument(
"--enable_mkldnn",
type=ast.literal_eval,
default=False,
help="Whether use mkldnn with CPU.")
parser.add_argument(
"--enable_mkldnn_bfloat16",
type=ast.literal_eval,
default=False,
help="Whether use mkldnn bfloat16 inference with CPU.")
parser.add_argument(
"--cpu_threads", type=int, default=1, help="Num of threads with CPU.")
parser.add_argument(
"--trt_min_shape", type=int, default=1, help="min_shape for TensorRT.")
parser.add_argument(
"--trt_max_shape",
type=int,
default=1280,
help="max_shape for TensorRT.")
parser.add_argument(
"--trt_opt_shape",
type=int,
default=640,
help="opt_shape for TensorRT.")
parser.add_argument(
"--trt_calib_mode",
type=bool,
default=False,
help="If the model is produced by TRT offline quantitative "
"calibration, trt_calib_mode need to set True.")
parser.add_argument(
'--save_images',
type=ast.literal_eval,
default=True,
help='Save visualization image results.')
parser.add_argument(
'--save_mot_txts',
action='store_true',
help='Save tracking results (txt).')
parser.add_argument(
'--save_mot_txt_per_img',
action='store_true',
help='Save tracking results (txt) for each image.')
parser.add_argument(
'--scaled',
type=bool,
default=False,
help="Whether coords after detector outputs are scaled, False in JDE YOLOv3 "
"True in general detector.")
parser.add_argument(
"--tracker_config", type=str, default=None, help=("tracker donfig"))
parser.add_argument(
"--reid_model_dir",
type=str,
default=None,
help=("Directory include:'model.pdiparams', 'model.pdmodel', "
"'infer_cfg.yml', created by tools/export_model.py."))
parser.add_argument(
"--reid_batch_size",
type=int,
default=50,
help="max batch_size for reid model inference.")
parser.add_argument(
'--use_dark',
type=ast.literal_eval,
default=True,
help='whether to use darkpose to get better keypoint position predict ')
parser.add_argument(
"--action_file",
type=str,
default=None,
help="Path of input file for action recognition.")
parser.add_argument(
"--window_size",
type=int,
default=50,
help="Temporal size of skeleton feature for action recognition.")
parser.add_argument(
"--random_pad",
type=ast.literal_eval,
default=False,
help="Whether do random padding for action recognition.")
parser.add_argument(
"--save_results",
action='store_true',
default=False,
help="Whether save detection result to file using coco format")
parser.add_argument(
'--use_coco_category',
action='store_true',
default=False,
help='Whether to use the coco format dictionary `clsid2catid`')
parser.add_argument(
"--slice_infer",
action='store_true',
help="Whether to slice the image and merge the inference results for small object detection."
)
parser.add_argument(
'--slice_size',
nargs='+',
type=int,
default=[640, 640],
help="Height of the sliced image.")
parser.add_argument(
"--overlap_ratio",
nargs='+',
type=float,
default=[0.25, 0.25],
help="Overlap height ratio of the sliced image.")
parser.add_argument(
"--combine_method",
type=str,
default='nms',
help="Combine method of the sliced images' detection results, choose in ['nms', 'nmm', 'concat']."
)
parser.add_argument(
"--match_threshold",
type=float,
default=0.6,
help="Combine method matching threshold.")
parser.add_argument(
"--match_metric",
type=str,
default='ios',
help="Combine method matching metric, choose in ['iou', 'ios'].")
parser.add_argument(
"--collect_trt_shape_info",
action='store_true',
default=False,
help="Whether to collect dynamic shape before using tensorrt.")
parser.add_argument(
"--tuned_trt_shape_file",
type=str,
default="shape_range_info.pbtxt",
help="Path of a dynamic shape file for tensorrt.")
parser.add_argument("--use_fd_format", action="store_true")
parser.add_argument(
"--task_type",
type=str,
default='Detection',
help="How to save the coco result, it only work with save_results==True. Optional inputs are Rotate or Detection, default is Detection."
)
return parser
class Times(object):
def __init__(self):
self.time = 0.
# start time
self.st = 0.
# end time
self.et = 0.
def start(self):
self.st = time.time()
def end(self, repeats=1, accumulative=True):
self.et = time.time()
if accumulative:
self.time += (self.et - self.st) / repeats
else:
self.time = (self.et - self.st) / repeats
def reset(self):
self.time = 0.
self.st = 0.
self.et = 0.
def value(self):
return round(self.time, 4)
class Timer(Times):
def __init__(self, with_tracker=False):
super(Timer, self).__init__()
self.with_tracker = with_tracker
self.preprocess_time_s = Times()
self.inference_time_s = Times()
self.postprocess_time_s = Times()
self.tracking_time_s = Times()
self.img_num = 0
def info(self, average=False):
pre_time = self.preprocess_time_s.value()
infer_time = self.inference_time_s.value()
post_time = self.postprocess_time_s.value()
track_time = self.tracking_time_s.value()
total_time = pre_time + infer_time + post_time
if self.with_tracker:
total_time = total_time + track_time
total_time = round(total_time, 4)
print("------------------ Inference Time Info ----------------------")
print("total_time(ms): {}, img_num: {}".format(total_time * 1000,
self.img_num))
preprocess_time = round(pre_time / max(1, self.img_num),
4) if average else pre_time
postprocess_time = round(post_time / max(1, self.img_num),
4) if average else post_time
inference_time = round(infer_time / max(1, self.img_num),
4) if average else infer_time
tracking_time = round(track_time / max(1, self.img_num),
4) if average else track_time
average_latency = total_time / max(1, self.img_num)
qps = 0
if total_time > 0:
qps = 1 / average_latency
print("average latency time(ms): {:.2f}, QPS: {:2f}".format(
average_latency * 1000, qps))
if self.with_tracker:
print(
"preprocess_time(ms): {:.2f}, inference_time(ms): {:.2f}, postprocess_time(ms): {:.2f}, tracking_time(ms): {:.2f}".
format(preprocess_time * 1000, inference_time * 1000,
postprocess_time * 1000, tracking_time * 1000))
else:
print(
"preprocess_time(ms): {:.2f}, inference_time(ms): {:.2f}, postprocess_time(ms): {:.2f}".
format(preprocess_time * 1000, inference_time * 1000,
postprocess_time * 1000))
def report(self, average=False):
dic = {}
pre_time = self.preprocess_time_s.value()
infer_time = self.inference_time_s.value()
post_time = self.postprocess_time_s.value()
track_time = self.tracking_time_s.value()
dic['preprocess_time_s'] = round(pre_time / max(1, self.img_num),
4) if average else pre_time
dic['inference_time_s'] = round(infer_time / max(1, self.img_num),
4) if average else infer_time
dic['postprocess_time_s'] = round(post_time / max(1, self.img_num),
4) if average else post_time
dic['img_num'] = self.img_num
total_time = pre_time + infer_time + post_time
if self.with_tracker:
dic['tracking_time_s'] = round(track_time / max(1, self.img_num),
4) if average else track_time
total_time = total_time + track_time
dic['total_time_s'] = round(total_time, 4)
return dic
def multiclass_nms(bboxs, num_classes, match_threshold=0.6, match_metric='iou'):
final_boxes = []
for c in range(num_classes):
idxs = bboxs[:, 0] == c
if np.count_nonzero(idxs) == 0: continue
r = nms(bboxs[idxs, 1:], match_threshold, match_metric)
final_boxes.append(np.concatenate([np.full((r.shape[0], 1), c), r], 1))
return final_boxes
def nms(dets, match_threshold=0.6, match_metric='iou'):
""" Apply NMS to avoid detecting too many overlapping bounding boxes.
Args:
dets: shape [N, 5], [score, x1, y1, x2, y2]
match_metric: 'iou' or 'ios'
match_threshold: overlap thresh for match metric.
"""
if dets.shape[0] == 0:
return dets[[], :]
scores = dets[:, 0]
x1 = dets[:, 1]
y1 = dets[:, 2]
x2 = dets[:, 3]
y2 = dets[:, 4]
areas = (x2 - x1 + 1) * (y2 - y1 + 1)
order = scores.argsort()[::-1]
ndets = dets.shape[0]
suppressed = np.zeros((ndets), dtype=np.int32)
for _i in range(ndets):
i = order[_i]
if suppressed[i] == 1:
continue
ix1 = x1[i]
iy1 = y1[i]
ix2 = x2[i]
iy2 = y2[i]
iarea = areas[i]
for _j in range(_i + 1, ndets):
j = order[_j]
if suppressed[j] == 1:
continue
xx1 = max(ix1, x1[j])
yy1 = max(iy1, y1[j])
xx2 = min(ix2, x2[j])
yy2 = min(iy2, y2[j])
w = max(0.0, xx2 - xx1 + 1)
h = max(0.0, yy2 - yy1 + 1)
inter = w * h
if match_metric == 'iou':
union = iarea + areas[j] - inter
match_value = inter / union
elif match_metric == 'ios':
smaller = min(iarea, areas[j])
match_value = inter / smaller
else:
raise ValueError()
if match_value >= match_threshold:
suppressed[j] = 1
keep = np.where(suppressed == 0)[0]
dets = dets[keep, :]
return dets
coco_clsid2catid = {
0: 1,
1: 2,
2: 3,
3: 4,
4: 5,
5: 6,
6: 7,
7: 8,
8: 9,
9: 10,
10: 11,
11: 13,
12: 14,
13: 15,
14: 16,
15: 17,
16: 18,
17: 19,
18: 20,
19: 21,
20: 22,
21: 23,
22: 24,
23: 25,
24: 27,
25: 28,
26: 31,
27: 32,
28: 33,
29: 34,
30: 35,
31: 36,
32: 37,
33: 38,
34: 39,
35: 40,
36: 41,
37: 42,
38: 43,
39: 44,
40: 46,
41: 47,
42: 48,
43: 49,
44: 50,
45: 51,
46: 52,
47: 53,
48: 54,
49: 55,
50: 56,
51: 57,
52: 58,
53: 59,
54: 60,
55: 61,
56: 62,
57: 63,
58: 64,
59: 65,
60: 67,
61: 70,
62: 72,
63: 73,
64: 74,
65: 75,
66: 76,
67: 77,
68: 78,
69: 79,
70: 80,
71: 81,
72: 82,
73: 84,
74: 85,
75: 86,
76: 87,
77: 88,
78: 89,
79: 90
}
def gaussian_radius(bbox_size, min_overlap):
height, width = bbox_size
a1 = 1
b1 = (height + width)
c1 = width * height * (1 - min_overlap) / (1 + min_overlap)
sq1 = np.sqrt(b1**2 - 4 * a1 * c1)
radius1 = (b1 + sq1) / (2 * a1)
a2 = 4
b2 = 2 * (height + width)
c2 = (1 - min_overlap) * width * height
sq2 = np.sqrt(b2**2 - 4 * a2 * c2)
radius2 = (b2 + sq2) / 2
a3 = 4 * min_overlap
b3 = -2 * min_overlap * (height + width)
c3 = (min_overlap - 1) * width * height
sq3 = np.sqrt(b3**2 - 4 * a3 * c3)
radius3 = (b3 + sq3) / 2
return min(radius1, radius2, radius3)
def gaussian2D(shape, sigma_x=1, sigma_y=1):
m, n = [(ss - 1.) / 2. for ss in shape]
y, x = np.ogrid[-m:m + 1, -n:n + 1]
h = np.exp(-(x * x / (2 * sigma_x * sigma_x) + y * y / (2 * sigma_y *
sigma_y)))
h[h < np.finfo(h.dtype).eps * h.max()] = 0
return h
def draw_umich_gaussian(heatmap, center, radius, k=1):
"""
draw_umich_gaussian, refer to https://github.com/xingyizhou/CenterNet/blob/master/src/lib/utils/image.py#L126
"""
diameter = 2 * radius + 1
gaussian = gaussian2D(
(diameter, diameter), sigma_x=diameter / 6, sigma_y=diameter / 6)
x, y = int(center[0]), int(center[1])
height, width = heatmap.shape[0:2]
left, right = min(x, radius), min(width - x, radius + 1)
top, bottom = min(y, radius), min(height - y, radius + 1)
masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right]
masked_gaussian = gaussian[radius - top:radius + bottom, radius - left:
radius + right]
if min(masked_gaussian.shape) > 0 and min(masked_heatmap.shape) > 0:
np.maximum(masked_heatmap, masked_gaussian * k, out=masked_heatmap)
return heatmap
def iou(box1, box2):
"""计算两个框的 IoU（交并比）"""
x1 = max(box1[0], box2[0])
y1 = max(box1[1], box2[1])
x2 = min(box1[2], box2[2])
y2 = min(box1[3], box2[3])
inter_area = max(0, x2 - x1) * max(0, y2 - y1)
box1_area = (box1[2] - box1[0]) * (box1[3] - box1[1])
box2_area = (box2[2] - box2[0]) * (box2[3] - box2[1])
union_area = box1_area + box2_area - inter_area
return inter_area / union_area if union_area != 0 else 0
def non_max_suppression(boxes, scores, iou_threshold):
"""非极大值抑制"""
if not boxes:
return []
indices = np.argsort(scores)[::-1]
selected_boxes = []
while len(indices) > 0:
current = indices[0]
selected_boxes.append(current)
remaining = indices[1:]
filtered_indices = []
for i in remaining:
if iou(boxes[current], boxes[i]) <= iou_threshold:
filtered_indices.append(i)
indices = np.array(filtered_indices)
return selected_boxes
def is_box_inside(inner, outer):
"""判断inner box是否完全在outer box内"""
return (outer[0] <= inner[0] and outer[1] <= inner[1] and
outer[2] >= inner[2] and outer[3] >= inner[3])
def boxes_overlap(box1: List[float], box2: List[float]) -> bool:
"""判断两个框是否有交集"""
x1_max = max(box1[0], box2[0])
y1_max = max(box1[1], box2[1])
x2_min = min(box1[2], box2[2])
y2_min = min(box1[3], box2[3])
return x1_max < x2_min and y1_max < y2_min
def merge_boxes(boxes: List[List[float]]) -> List[float]:
x1 = min(box[0] for box in boxes)
y1 = min(box[1] for box in boxes)
x2 = max(box[2] for box in boxes)
y2 = max(box[3] for box in boxes)
return [x1, y1, x2, y2]
def merge_text_and_title_boxes(data: List[LayoutBox], merged_labels: Tuple[int]) -> List[LayoutBox]:
text_title_boxes = [(i, box) for i, box in enumerate(data) if box.clsid in merged_labels]
other_boxes = [box.pos for box in data if box.clsid in (2, 4, 5)]
text_title_boxes.sort(key=lambda x: x[1].pos[1]) # sort by y1
merged = []
skip_indices = set()
i = 0
while i < len(text_title_boxes):
if i in skip_indices:
i += 1
continue
current_group = [text_title_boxes[i][1].pos]
group_confidences = [text_title_boxes[i][1].confidence]
j = i + 1
while j < len(text_title_boxes):
candidate_box = text_title_boxes[j][1].pos
tentative_merge = merge_boxes(current_group + [candidate_box])
has_intruder = any(boxes_overlap(other, tentative_merge) for other in other_boxes)
if has_intruder:
break
else:
current_group.append(candidate_box)
group_confidences.append(text_title_boxes[j][1].confidence)
skip_indices.add(j)
j += 1
if len(current_group) > 1:
merged_box = LayoutBox(0, merge_boxes(current_group), max(group_confidences))
merged.append(merged_box)
else:
idx = text_title_boxes[i][0]
merged.append(data[idx])
i += 1
remaining = [data[i] for i in range(len(data)) if i not in skip_indices and data[i].clsid not in merged_labels]
return merged + remaining