File size: 9,568 Bytes
3ec4928 |
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 |
"""
Ref-YoutubeVOS data loader
"""
from pathlib import Path
import torch
from torch.autograd.grad_mode import F
from torch.utils.data import Dataset
import datasets.transforms_video as T
import os
from PIL import Image
import json
import numpy as np
import random
from datasets.categories import ytvos_category_dict as category_dict
class YTVOSDataset(Dataset):
"""
A dataset class for the Refer-Youtube-VOS dataset which was first introduced in the paper:
"URVOS: Unified Referring Video Object Segmentation Network with a Large-Scale Benchmark"
(see https://link.springer.com/content/pdf/10.1007/978-3-030-58555-6_13.pdf).
The original release of the dataset contained both 'first-frame' and 'full-video' expressions. However, the first
dataset is not publicly available anymore as now only the harder 'full-video' subset is available to download
through the Youtube-VOS referring video object segmentation competition page at:
https://competitions.codalab.org/competitions/29139
Furthermore, for the competition the subset's original validation set, which consists of 507 videos, was split into
two competition 'validation' & 'test' subsets, consisting of 202 and 305 videos respectively. Evaluation can
currently only be done on the competition 'validation' subset using the competition's server, as
annotations were publicly released only for the 'train' subset of the competition.
"""
def __init__(self, img_folder: Path, ann_file: Path, transforms, return_masks: bool,
num_frames: int, max_skip: int):
self.img_folder = img_folder
self.ann_file = ann_file
self._transforms = transforms
self.return_masks = return_masks # not used
self.num_frames = num_frames
self.max_skip = max_skip
# create video meta data
self.prepare_metas()
print('\n video num: ', len(self.videos), ' clip num: ', len(self.metas))
print('\n')
def prepare_metas(self):
# read object information
with open(os.path.join(str(self.img_folder), 'meta.json'), 'r') as f:
subset_metas_by_video = json.load(f)['videos']
# read expression data
with open(str(self.ann_file), 'r') as f:
subset_expressions_by_video = json.load(f)['videos']
self.videos = list(subset_expressions_by_video.keys())
self.metas = []
for vid in self.videos:
vid_meta = subset_metas_by_video[vid]
vid_data = subset_expressions_by_video[vid]
vid_frames = sorted(vid_data['frames'])
vid_len = len(vid_frames)
print(vid_meta)
for exp_id, exp_dict in vid_data['expressions'].items():
print(exp_dict)
# Exclude start_idx (0, 1) and end_idx (vid_len-1, vid_len-2)
start_idx , end_idx = 2, vid_len-2
bin_size = (end_idx - start_idx) // 4
bins = []
for i in range(4):
bin_start = start_idx + i * bin_size
bin_end = bin_start + bin_size if i < 3 else end_idx
bins.append((bin_start, bin_end))
meta = {
'video': vid,
'exp': exp_dict['exp'],
'obj_id': int(exp_dict['obj_id']),
'frames': vid_frames,
'bins': bins,
'category': vid_meta['objects'][exp_dict['obj_id']]['category']
}
self.metas.append(meta)
@staticmethod
def bounding_box(img):
rows = np.any(img, axis=1)
cols = np.any(img, axis=0)
rmin, rmax = np.where(rows)[0][[0, -1]]
cmin, cmax = np.where(cols)[0][[0, -1]]
return rmin, rmax, cmin, cmax # y1, y2, x1, x2
def __len__(self):
return len(self.metas)
def __getitem__(self, idx):
instance_check = False
while not instance_check:
meta = self.metas[idx] # dict
video, exp, obj_id, category, frames, bins = \
meta['video'], meta['exp'], meta['obj_id'], meta['category'], meta['frames'], meta['bins']
# clean up the caption
exp = " ".join(exp.lower().split())
category_id = category_dict[category]
vid_len = len(frames)
# num_frames = self.num_frames
# Random sample one frame from each bin
sample_indx = []
for start_idx, end_idx in bins:
sample_indx.append(random.randint(start_idx, end_idx - 1))
sample_indx.sort() # Ensure indices are in order
# read frames and masks
imgs, labels, boxes, masks, valid = [], [], [], [], []
for frame_indx in sample_indx:
frame_name = frames[frame_indx]
img_path = os.path.join(str(self.img_folder), 'JPEGImages', video, frame_name + '.jpg')
mask_path = os.path.join(str(self.img_folder), 'Annotations', video, frame_name + '.png')
img = Image.open(img_path).convert('RGB')
mask = Image.open(mask_path).convert('P')
# create the target
label = torch.tensor(category_id)
mask = np.array(mask)
mask = (mask==obj_id).astype(np.float32) # 0,1 binary
if (mask > 0).any():
y1, y2, x1, x2 = self.bounding_box(mask)
box = torch.tensor([x1, y1, x2, y2]).to(torch.float)
valid.append(1)
else: # some frame didn't contain the instance
box = torch.tensor([0, 0, 0, 0]).to(torch.float)
valid.append(0)
mask = torch.from_numpy(mask)
# append
imgs.append(img)
labels.append(label)
masks.append(mask)
boxes.append(box)
# transform
w, h = img.size
labels = torch.stack(labels, dim=0)
boxes = torch.stack(boxes, dim=0)
boxes[:, 0::2].clamp_(min=0, max=w)
boxes[:, 1::2].clamp_(min=0, max=h)
masks = torch.stack(masks, dim=0)
target = {
'frames_idx': torch.tensor(sample_indx), # [T,]
'labels': labels, # [T,]
'boxes': boxes, # [T, 4], xyxy
'masks': masks, # [T, H, W]
'valid': torch.tensor(valid), # [T,]
'caption': exp,
'orig_size': torch.as_tensor([int(h), int(w)]),
'size': torch.as_tensor([int(h), int(w)])
}
# "boxes" normalize to [0, 1] and transform from xyxy to cxcywh in self._transform
if self._transforms:
imgs, target = self._transforms(imgs, target)
imgs = torch.stack(imgs, dim=0) # [T, 3, H, W]
else:
imgs = np.array(imgs)
imgs = torch.tensor(imgs.transpose(0, 3, 1, 2))
# FIXME: handle "valid", since some box may be removed due to random crop
if torch.any(target['valid'] == 1): # at leatst one instance
instance_check = True
else:
idx = random.randint(0, self.__len__() - 1)
return imgs, target
def make_coco_transforms(image_set, max_size=640):
normalize = T.Compose([
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [288, 320, 352, 392, 416, 448, 480, 512]
if image_set == 'train':
return T.Compose([
T.RandomHorizontalFlip(),
T.PhotometricDistort(),
T.RandomSelect(
T.Compose([
T.RandomResize(scales, max_size=max_size),
T.Check(),
]),
T.Compose([
T.RandomResize([400, 500, 600]),
T.RandomSizeCrop(384, 600),
T.RandomResize(scales, max_size=max_size),
T.Check(),
])
),
normalize,
])
# we do not use the 'val' set since the annotations are inaccessible
if image_set == 'val':
return T.Compose([
T.RandomResize([360], max_size=640),
normalize,
])
raise ValueError(f'unknown {image_set}')
def build(image_set, args):
root = Path(args.ytvos_path)
assert root.exists(), f'provided YTVOS path {root} does not exist'
PATHS = {
"train": (root / "train", root / "meta_expressions" / "train" / "meta_expressions.json"),
"val": (root / "valid", root / "meta_expressions" / "valid" / "meta_expressions.json"), # not used actually
}
img_folder, ann_file = PATHS[image_set]
# dataset = YTVOSDataset(img_folder, ann_file, transforms=make_coco_transforms(image_set, max_size=args.max_size), return_masks=args.masks,
# num_frames=args.num_frames, max_skip=args.max_skip)
dataset = YTVOSDataset(img_folder, ann_file, transforms=None, return_masks=args.masks,
num_frames=args.num_frames, max_skip=args.max_skip)
return dataset
|