File size: 11,759 Bytes
483de47 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 |
# ------------------------------------------------------------------------
# Copyright (c) 2022 megvii-research. All Rights Reserved.
# ------------------------------------------------------------------------
# Modified from Deformable DETR (https://github.com/fundamentalvision/Deformable-DETR)
# Copyright (c) 2020 SenseTime. All Rights Reserved.
# ------------------------------------------------------------------------
# Modified from DETR (https://github.com/facebookresearch/detr)
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
# ------------------------------------------------------------------------
"""
MOT dataset which returns image_id for evaluation.
"""
from pathlib import Path
import cv2
import numpy as np
import torch
import torch.utils.data
import os.path as osp
from PIL import Image, ImageDraw
import copy
import datasets.transforms as T
from models.structures import Instances
class DetMOTDetection:
def __init__(self, args, data_txt_path: str, seqs_folder, dataset2transform):
self.args = args
self.dataset2transform = dataset2transform
self.num_frames_per_batch = max(args.sampler_lengths)
self.sample_mode = args.sample_mode
self.sample_interval = args.sample_interval
self.vis = args.vis
self.video_dict = {}
with open(data_txt_path, 'r') as file:
self.img_files = file.readlines()
self.img_files = [osp.join(seqs_folder, x.strip()) for x in self.img_files]
self.img_files = list(filter(lambda x: len(x) > 0, self.img_files))
self.label_files = [(x.replace('images', 'labels_with_ids').replace('.png', '.txt').replace('.jpg', '.txt'))
for x in self.img_files]
# The number of images per sample: 1 + (num_frames - 1) * interval.
# The number of valid samples: num_images - num_image_per_sample + 1.
self.item_num = len(self.img_files) - (self.num_frames_per_batch - 1) * self.sample_interval
self._register_videos()
# video sampler.
self.sampler_steps: list = args.sampler_steps
self.lengths: list = args.sampler_lengths
print("sampler_steps={} lenghts={}".format(self.sampler_steps, self.lengths))
if self.sampler_steps is not None and len(self.sampler_steps) > 0:
# Enable sampling length adjustment.
assert len(self.lengths) > 0
assert len(self.lengths) == len(self.sampler_steps) + 1
for i in range(len(self.sampler_steps) - 1):
assert self.sampler_steps[i] < self.sampler_steps[i + 1]
self.item_num = len(self.img_files) - (self.lengths[-1] - 1) * self.sample_interval
self.period_idx = 0
self.num_frames_per_batch = self.lengths[0]
self.current_epoch = 0
def _register_videos(self):
for label_name in self.label_files:
video_name = '/'.join(label_name.split('/')[:-1])
if video_name not in self.video_dict:
print("register {}-th video: {} ".format(len(self.video_dict) + 1, video_name))
self.video_dict[video_name] = len(self.video_dict)
# assert len(self.video_dict) <= 300
def set_epoch(self, epoch):
self.current_epoch = epoch
if self.sampler_steps is None or len(self.sampler_steps) == 0:
# fixed sampling length.
return
for i in range(len(self.sampler_steps)):
if epoch >= self.sampler_steps[i]:
self.period_idx = i + 1
print("set epoch: epoch {} period_idx={}".format(epoch, self.period_idx))
self.num_frames_per_batch = self.lengths[self.period_idx]
def step_epoch(self):
# one epoch finishes.
print("Dataset: epoch {} finishes".format(self.current_epoch))
self.set_epoch(self.current_epoch + 1)
@staticmethod
def _targets_to_instances(targets: dict, img_shape) -> Instances:
gt_instances = Instances(tuple(img_shape))
gt_instances.boxes = targets['boxes']
gt_instances.labels = targets['labels']
gt_instances.obj_ids = targets['obj_ids']
gt_instances.area = targets['area']
return gt_instances
def _pre_single_frame(self, idx: int):
img_path = self.img_files[idx]
label_path = self.label_files[idx]
if 'crowdhuman' in img_path:
img_path = img_path.replace('.jpg', '.png')
img = Image.open(img_path)
targets = {}
w, h = img._size
assert w > 0 and h > 0, "invalid image {} with shape {} {}".format(img_path, w, h)
if osp.isfile(label_path):
labels0 = np.loadtxt(label_path, dtype=np.float32).reshape(-1, 6)
# normalized cewh to pixel xyxy format
labels = labels0.copy()
labels[:, 2] = w * (labels0[:, 2] - labels0[:, 4] / 2)
labels[:, 3] = h * (labels0[:, 3] - labels0[:, 5] / 2)
labels[:, 4] = w * (labels0[:, 2] + labels0[:, 4] / 2)
labels[:, 5] = h * (labels0[:, 3] + labels0[:, 5] / 2)
else:
raise ValueError('invalid label path: {}'.format(label_path))
video_name = '/'.join(label_path.split('/')[:-1])
obj_idx_offset = self.video_dict[video_name] * 1000000 # 1000000 unique ids is enough for a video.
if 'crowdhuman' in img_path:
targets['dataset'] = 'CrowdHuman'
elif 'MOT17' in img_path:
targets['dataset'] = 'MOT17'
else:
raise NotImplementedError()
targets['boxes'] = []
targets['area'] = []
targets['iscrowd'] = []
targets['labels'] = []
targets['obj_ids'] = []
targets['image_id'] = torch.as_tensor(idx)
targets['size'] = torch.as_tensor([h, w])
targets['orig_size'] = torch.as_tensor([h, w])
for label in labels:
targets['boxes'].append(label[2:6].tolist())
targets['area'].append(label[4] * label[5])
targets['iscrowd'].append(0)
targets['labels'].append(0)
obj_id = label[1] + obj_idx_offset if label[1] >= 0 else label[1]
targets['obj_ids'].append(obj_id) # relative id
targets['area'] = torch.as_tensor(targets['area'])
targets['iscrowd'] = torch.as_tensor(targets['iscrowd'])
targets['labels'] = torch.as_tensor(targets['labels'])
targets['obj_ids'] = torch.as_tensor(targets['obj_ids'])
targets['boxes'] = torch.as_tensor(targets['boxes'], dtype=torch.float32).reshape(-1, 4)
return img, targets
def _get_sample_range(self, start_idx):
# take default sampling method for normal dataset.
assert self.sample_mode in ['fixed_interval', 'random_interval'], 'invalid sample mode: {}'.format(self.sample_mode)
if self.sample_mode == 'fixed_interval':
sample_interval = self.sample_interval
elif self.sample_mode == 'random_interval':
sample_interval = np.random.randint(1, self.sample_interval + 1)
default_range = start_idx, start_idx + (self.num_frames_per_batch - 1) * sample_interval + 1, sample_interval
return default_range
def pre_continuous_frames(self, start, end, interval=1):
targets = []
images = []
for i in range(start, end, interval):
img_i, targets_i = self._pre_single_frame(i)
images.append(img_i)
targets.append(targets_i)
return images, targets
def __getitem__(self, idx):
sample_start, sample_end, sample_interval = self._get_sample_range(idx)
images, targets = self.pre_continuous_frames(sample_start, sample_end, sample_interval)
data = {}
dataset_name = targets[0]['dataset']
transform = self.dataset2transform[dataset_name]
if transform is not None:
images, targets = transform(images, targets)
gt_instances = []
for img_i, targets_i in zip(images, targets):
gt_instances_i = self._targets_to_instances(targets_i, img_i.shape[1:3])
gt_instances.append(gt_instances_i)
data.update({
'imgs': images,
'gt_instances': gt_instances,
})
if self.args.vis:
data['ori_img'] = [target_i['ori_img'] for target_i in targets]
return data
def __len__(self):
return self.item_num
class DetMOTDetectionValidation(DetMOTDetection):
def __init__(self, args, seqs_folder, dataset2transform):
args.data_txt_path = args.val_data_txt_path
super().__init__(args, seqs_folder, dataset2transform)
def make_transforms_for_mot17(image_set, args=None):
normalize = T.MotCompose([
T.MotToTensor(),
T.MotNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [608, 640, 672, 704, 736, 768, 800, 832, 864, 896, 928, 960, 992]
if image_set == 'train':
return T.MotCompose([
T.MotRandomHorizontalFlip(),
T.MotRandomSelect(
T.MotRandomResize(scales, max_size=1536),
T.MotCompose([
T.MotRandomResize([400, 500, 600]),
T.FixedMotRandomCrop(384, 600),
T.MotRandomResize(scales, max_size=1536),
])
),
normalize,
])
if image_set == 'val':
return T.MotCompose([
T.MotRandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def make_transforms_for_crowdhuman(image_set, args=None):
normalize = T.MotCompose([
T.MotToTensor(),
T.MotNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [608, 640, 672, 704, 736, 768, 800, 832, 864, 896, 928, 960, 992]
if image_set == 'train':
return T.MotCompose([
T.MotRandomHorizontalFlip(),
T.FixedMotRandomShift(bs=1),
T.MotRandomSelect(
T.MotRandomResize(scales, max_size=1536),
T.MotCompose([
T.MotRandomResize([400, 500, 600]),
T.FixedMotRandomCrop(384, 600),
T.MotRandomResize(scales, max_size=1536),
])
),
normalize,
])
if image_set == 'val':
return T.MotCompose([
T.MotRandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def build_dataset2transform(args, image_set):
mot17_train = make_transforms_for_mot17('train', args)
mot17_test = make_transforms_for_mot17('val', args)
crowdhuman_train = make_transforms_for_crowdhuman('train', args)
dataset2transform_train = {'MOT17': mot17_train, 'CrowdHuman': crowdhuman_train}
dataset2transform_val = {'MOT17': mot17_test, 'CrowdHuman': mot17_test}
if image_set == 'train':
return dataset2transform_train
elif image_set == 'val':
return dataset2transform_val
else:
raise NotImplementedError()
def build(image_set, args):
root = Path(args.mot_path)
assert root.exists(), f'provided MOT path {root} does not exist'
dataset2transform = build_dataset2transform(args, image_set)
if image_set == 'train':
data_txt_path = args.data_txt_path_train
dataset = DetMOTDetection(args, data_txt_path=data_txt_path, seqs_folder=root, dataset2transform=dataset2transform)
if image_set == 'val':
data_txt_path = args.data_txt_path_val
dataset = DetMOTDetection(args, data_txt_path=data_txt_path, seqs_folder=root, dataset2transform=dataset2transform)
return dataset
|