File size: 10,912 Bytes
864ec44 |
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 |
# Last modified: 2024-04-30
#
# Copyright 2023 Bingxin Ke, ETH Zurich. 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.
# --------------------------------------------------------------------------
# If you find this code useful, we kindly ask you to cite our paper in your work.
# Please find bibtex at: https://github.com/prs-eth/Marigold#-citation
# If you use or adapt this code, please attribute to https://github.com/prs-eth/marigold.
# More information about the method can be found at https://marigoldmonodepth.github.io
# --------------------------------------------------------------------------
import glob
import io
import json
import os
import pdb
import random
import tarfile
from enum import Enum
from typing import Union
import numpy as np
import torch
from PIL import Image
from torch.utils.data import Dataset
from torchvision.transforms import InterpolationMode, Resize, CenterCrop
import torchvision.transforms as transforms
from transformers import CLIPTextModel, CLIPTokenizer
from src.util.depth_transform import DepthNormalizerBase
import random
from src.dataset.eval_base_dataset import DatasetMode, DepthFileNameMode
def read_image_from_tar(tar_obj, img_rel_path):
image = tar_obj.extractfile("./" + img_rel_path)
image = image.read()
image = Image.open(io.BytesIO(image))
class BaseDepthDataset(Dataset):
def __init__(
self,
mode: DatasetMode,
filename_ls_path: str,
dataset_dir: str,
disp_name: str,
min_depth: float,
max_depth: float,
has_filled_depth: bool,
name_mode: DepthFileNameMode,
depth_transform: Union[DepthNormalizerBase, None] = None,
tokenizer: CLIPTokenizer = None,
augmentation_args: dict = None,
resize_to_hw=None,
move_invalid_to_far_plane: bool = True,
rgb_transform=lambda x: x / 255.0 * 2 - 1, # [0, 255] -> [-1, 1],
**kwargs,
) -> None:
super().__init__()
self.mode = mode
# dataset info
self.filename_ls_path = filename_ls_path
self.disp_name = disp_name
self.has_filled_depth = has_filled_depth
self.name_mode: DepthFileNameMode = name_mode
self.min_depth = min_depth
self.max_depth = max_depth
# training arguments
self.depth_transform: DepthNormalizerBase = depth_transform
self.augm_args = augmentation_args
self.resize_to_hw = resize_to_hw
self.rgb_transform = rgb_transform
self.move_invalid_to_far_plane = move_invalid_to_far_plane
self.tokenizer = tokenizer
# Load filenames
self.filenames = []
filename_paths = glob.glob(self.filename_ls_path)
for path in filename_paths:
with open(path, "r") as f:
self.filenames += json.load(f)
# Tar dataset
self.tar_obj = None
self.is_tar = (
True
if os.path.isfile(dataset_dir) and tarfile.is_tarfile(dataset_dir)
else False
)
def __len__(self):
return len(self.filenames)
def __getitem__(self, index):
rasters, other = self._get_data_item(index)
if DatasetMode.TRAIN == self.mode:
rasters = self._training_preprocess(rasters)
# merge
outputs = rasters
outputs.update(other)
return outputs
def _get_data_item(self, index):
rgb_path = self.filenames[index]['rgb_path']
depth_path = self.filenames[index]['depth_path']
mask_path = None
if 'valid_mask' in self.filenames[index]:
mask_path = self.filenames[index]['valid_mask']
if self.filenames[index]['caption'] is not None:
coca_caption = self.filenames[index]['caption']['coca_caption']
spatial_caption = self.filenames[index]['caption']['spatial_caption']
empty_caption = ''
caption_choices = [coca_caption, spatial_caption, empty_caption]
probabilities = [0.4, 0.4, 0.2]
caption = random.choices(caption_choices, probabilities)[0]
else:
caption = ''
rasters = {}
# RGB data
rasters.update(self._load_rgb_data(rgb_path))
# Depth data
if DatasetMode.RGB_ONLY != self.mode and depth_path is not None:
# load data
depth_data = self._load_depth_data(depth_path)
rasters.update(depth_data)
# valid mask
if mask_path is not None:
valid_mask_raw = Image.open(mask_path)
valid_mask_filled = Image.open(mask_path)
rasters["valid_mask_raw"] = torch.from_numpy(np.asarray(valid_mask_raw)).unsqueeze(0).bool()
rasters["valid_mask_filled"] = torch.from_numpy(np.asarray(valid_mask_filled)).unsqueeze(0).bool()
else:
rasters["valid_mask_raw"] = self._get_valid_mask(
rasters["depth_raw_linear"]
).clone()
rasters["valid_mask_filled"] = self._get_valid_mask(
rasters["depth_filled_linear"]
).clone()
other = {"index": index, "rgb_path": rgb_path, 'text': caption}
if self.resize_to_hw is not None:
resize_transform = transforms.Compose([
Resize(size=max(self.resize_to_hw), interpolation=InterpolationMode.NEAREST_EXACT),
CenterCrop(size=self.resize_to_hw)])
rasters = {k: resize_transform(v) for k, v in rasters.items()}
return rasters, other
def _load_rgb_data(self, rgb_path):
# Read RGB data
rgb = self._read_rgb_file(rgb_path)
rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
outputs = {
"rgb_int": torch.from_numpy(rgb).int(),
"rgb_norm": torch.from_numpy(rgb_norm).float(),
}
return outputs
def _load_depth_data(self, depth_path, filled_rel_path=None):
# Read depth data
outputs = {}
depth_raw = self._read_depth_file(depth_path).squeeze()
depth_raw_linear = torch.from_numpy(depth_raw.copy()).float().unsqueeze(0) # [1, H, W]
outputs["depth_raw_linear"] = depth_raw_linear.clone()
if self.has_filled_depth:
depth_filled = self._read_depth_file(filled_rel_path).squeeze()
depth_filled_linear = torch.from_numpy(depth_filled).float().unsqueeze(0)
outputs["depth_filled_linear"] = depth_filled_linear
else:
outputs["depth_filled_linear"] = depth_raw_linear.clone()
return outputs
def _get_data_path(self, index):
filename_line = self.filenames[index]
# Get data path
rgb_rel_path = filename_line[0]
depth_rel_path, text_rel_path = None, None
if DatasetMode.RGB_ONLY != self.mode:
depth_rel_path = filename_line[1]
if len(filename_line) > 2:
text_rel_path = filename_line[2]
return rgb_rel_path, depth_rel_path, text_rel_path
def _read_image(self, img_path) -> np.ndarray:
image_to_read = img_path
image = Image.open(image_to_read) # [H, W, rgb]
image = np.asarray(image)
return image
def _read_rgb_file(self, path) -> np.ndarray:
rgb = self._read_image(path)
rgb = np.transpose(rgb, (2, 0, 1)).astype(int) # [rgb, H, W]
return rgb
def _read_depth_file(self, path):
depth_in = self._read_image(path)
# Replace code below to decode depth according to dataset definition
depth_decoded = depth_in
return depth_decoded
def _get_valid_mask(self, depth: torch.Tensor):
valid_mask = torch.logical_and(
(depth > self.min_depth), (depth < self.max_depth)
).bool()
return valid_mask
def _training_preprocess(self, rasters):
# Augmentation
if self.augm_args is not None:
rasters = self._augment_data(rasters)
# Normalization
# rasters["depth_raw_norm"] = rasters["depth_raw_linear"] / 255.0 * 2.0 - 1.0
# rasters["depth_filled_norm"] = rasters["depth_filled_linear"] / 255.0 * 2.0 - 1.0
rasters["depth_raw_norm"] = self.depth_transform(
rasters["depth_raw_linear"], rasters["valid_mask_raw"]
).clone()
rasters["depth_filled_norm"] = self.depth_transform(
rasters["depth_filled_linear"], rasters["valid_mask_filled"]
).clone()
# Set invalid pixel to far plane
if self.move_invalid_to_far_plane:
if self.depth_transform.far_plane_at_max:
rasters["depth_filled_norm"][~rasters["valid_mask_filled"]] = (
self.depth_transform.norm_max
)
else:
rasters["depth_filled_norm"][~rasters["valid_mask_filled"]] = (
self.depth_transform.norm_min
)
# Resize
if self.resize_to_hw is not None:
resize_transform = transforms.Compose([
Resize(size=max(self.resize_to_hw), interpolation=InterpolationMode.NEAREST_EXACT),
CenterCrop(size=self.resize_to_hw)])
rasters = {k: resize_transform(v) for k, v in rasters.items()}
return rasters
def _augment_data(self, rasters_dict):
# lr flipping
lr_flip_p = self.augm_args.lr_flip_p
if random.random() < lr_flip_p:
rasters_dict = {k: v.flip(-1) for k, v in rasters_dict.items()}
return rasters_dict
def __del__(self):
if hasattr(self, "tar_obj") and self.tar_obj is not None:
self.tar_obj.close()
self.tar_obj = None
def get_pred_name(rgb_basename, name_mode, suffix=".png"):
if DepthFileNameMode.rgb_id == name_mode:
pred_basename = "pred_" + rgb_basename.split("_")[1]
elif DepthFileNameMode.i_d_rgb == name_mode:
pred_basename = rgb_basename.replace("_rgb.", "_pred.")
elif DepthFileNameMode.id == name_mode:
pred_basename = "pred_" + rgb_basename
elif DepthFileNameMode.rgb_i_d == name_mode:
pred_basename = "pred_" + "_".join(rgb_basename.split("_")[1:])
else:
raise NotImplementedError
# change suffix
pred_basename = os.path.splitext(pred_basename)[0] + suffix
return pred_basename
|