Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,354 Bytes
9e15541 |
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 |
import os
import pickle
import time
from pathlib import Path
from typing import Optional
import cv2
import numpy as np
import torch
from torch.utils.data import Dataset
from torchvision.transforms import ColorJitter
from scenedino.common.augmentation import get_color_aug_fn
class RealEstate10kDataset(Dataset):
def __init__(
self,
data_path: str,
split_path: Optional[str] = None,
target_image_size=(256, 384),
frame_count=2,
dilation=1,
keyframe_offset=0,
color_aug=False,
):
self.data_path = data_path
self.split_path = split_path
self.target_image_size = target_image_size
self.frame_count = frame_count
self.dilation = dilation
self.keyframe_offset = keyframe_offset
self.color_aug = color_aug
if self.split_path is None:
self.split = "train"
else:
self.split = "test"
self._seq_data = self._load_seq_data(self.data_path, self.split)
self._seq_keys = list(self._seq_data.keys())
if isinstance(self.dilation, int):
self._left_offset = (
(self.frame_count - 1) // 2 + self.keyframe_offset
) * self.dilation
dilation = self.dilation
else:
self._left_offset = 0
dilation = 0
if self.split == "train":
self._key_id_pairs = self._full_index(
self._seq_keys,
self._seq_data,
self._left_offset,
(self.frame_count - 1) * dilation,
)
else:
self._key_id_pairs = self._load_index(split_path)
self._skip = 0
self.length = len(self._key_id_pairs)
@staticmethod
def _load_seq_data(data_path, split):
file_path = Path(data_path) / f"{split}.pickle"
with open(file_path, "rb") as f:
seq_data = pickle.load(f)
return seq_data
@staticmethod
def _full_index(seq_keys, seq_data, left_offset, extra_frames):
key_id_pairs = []
for k in seq_keys:
seq_len = len(seq_data[k]["timestamps"])
seq_key_id_pairs = [
(k, i + left_offset) for i in range(seq_len - extra_frames)
]
key_id_pairs += seq_key_id_pairs
return key_id_pairs
@staticmethod
def _load_index(index_path):
def get_key_id(s):
parts = s.split(" ")
key = parts[0]
id = int(parts[1])
return key, id
with open(index_path, "r") as f:
lines = f.readlines()
key_id_pairs = list(map(get_key_id, lines))
return key_id_pairs
def load_images(self, key, ids):
imgs = []
for id in ids:
timestamp = self._seq_data[key]["timestamps"][id]
img = (
cv2.cvtColor(
cv2.imread(
os.path.join(
self.data_path,
"frames",
self.split,
key,
f"{timestamp}.jpg",
)
),
cv2.COLOR_BGR2RGB,
).astype(np.float32)
/ 255
)
imgs += [img]
return imgs
def process_img(self, img: np.array, color_aug_fn=None):
if self.target_image_size:
img = cv2.resize(
img,
(self.target_image_size[1], self.target_image_size[0]),
interpolation=cv2.INTER_LINEAR,
)
img = np.transpose(img, (2, 0, 1))
if color_aug_fn is not None:
img = color_aug_fn(torch.tensor(img))
img = img * 2 - 1
return img
@staticmethod
def process_pose(pose):
pose = np.concatenate(
(pose.astype(np.float32), np.array([[0, 0, 0, 1]], dtype=np.float32)),
axis=0,
)
pose = np.linalg.inv(pose)
return pose
@staticmethod
def process_projs(proj):
K = np.eye(3, dtype=np.float32)
K[0, 0] = 2 * proj[0]
K[1, 1] = 2 * proj[1]
K[0, 2] = 2 * proj[2] - 1
K[1, 2] = 2 * proj[3] - 1
return K
def __getitem__(self, index: int):
_start_time = time.time()
if index >= self.length:
raise IndexError()
if self._skip != 0:
index += self._skip
if self.color_aug:
color_aug_fn = get_color_aug_fn(
ColorJitter.get_params(
brightness=(0.8, 1.2),
contrast=(0.8, 1.2),
saturation=(0.8, 1.2),
hue=(-0.1, 0.1),
)
)
else:
color_aug_fn = None
key, index = self._key_id_pairs[index]
seq_len = len(self._seq_data[key]["timestamps"])
if self.dilation == "random":
dilation = torch.randint(1, 30, (1,)).item()
left_offset = self._left_offset
if self.frame_count > 2:
left_offset = dilation * (self.frame_count // 2)
else:
dilation = self.dilation
left_offset = self._left_offset
ids = [index] + [
max(min(i, seq_len - 1), 0)
for i in range(
index - left_offset,
index - left_offset + self.frame_count * dilation,
dilation,
)
if i != index
]
imgs = self.load_images(key, ids)
imgs = [self.process_img(img, color_aug_fn=color_aug_fn) for img in imgs]
# These poses are camera to world !!
poses = [self.process_pose(self._seq_data[key]["poses"][i, :, :]) for i in ids]
projs = [
self.process_projs(self._seq_data[key]["intrinsics"][i, :]) for i in ids
]
depths = [np.ones_like(imgs[0][:1, :, :])]
_proc_time = np.array(time.time() - _start_time)
data = {
"imgs": imgs,
"projs": projs,
"poses": poses,
"depths": depths,
"t__get_item__": np.array([_proc_time]),
}
return data
def __len__(self) -> int:
return self.length
|