File size: 6,938 Bytes
17cd746 |
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 |
#
# Toyota Motor Europe NV/SA and its affiliated companies retain all intellectual
# property and proprietary rights in and to this software and related documentation.
# Any commercial use, reproduction, disclosure or distribution of this software and
# related documentation without an express license agreement from Toyota Motor Europe NV/SA
# is strictly prohibited.
#
import json
import numpy as np
import torch
from vhap.data.video_dataset import VideoDataset
from vhap.config.nersemble import NersembleDataConfig
from vhap.util import camera
from vhap.util.log import get_logger
logger = get_logger(__name__)
class NeRSembleDataset(VideoDataset):
def __init__(
self,
cfg: NersembleDataConfig,
img_to_tensor: bool = False,
batchify_all_views: bool = False,
):
"""
Args:
root_folder: Path to dataset with the following directory layout
<root_folder>/
|---camera_params/
| |---<subject>/
| |---camera_params.json
|
|---color_correction/
| |---<subject>/
| |---<camera_id>.npy
|
|---<subject>/
|---<sequence>/
|---images/
| |---cam_<camera_id>_<timestep_id>.jpg
|
|---alpha_maps/
| |---cam_<camera_id>_<timestep_id>.png
|
|---landmark2d/
|---face-alignment/
| |---<camera_id>.npz
|
|---STAR/
|---<camera_id>.npz
"""
self.cfg = cfg
assert cfg.subject != "", "Please specify the subject name"
super().__init__(
cfg=cfg,
img_to_tensor=img_to_tensor,
batchify_all_views=batchify_all_views,
)
def match_sequences(self):
logger.info(f"Subject: {self.cfg.subject}, sequence: {self.cfg.sequence}")
return list(filter(lambda x: x.is_dir(), (self.cfg.root_folder / self.cfg.subject).glob(f"{self.cfg.sequence}*")))
def define_properties(self):
super().define_properties()
self.properties['rgb']['cam_id_prefix'] = "cam_"
self.properties['alpha_map']['cam_id_prefix'] = "cam_"
def load_camera_params(self):
load_path = self.cfg.root_folder / "camera_params" / self.cfg.subject / "camera_params.json"
assert load_path.exists()
param = json.load(open(load_path))
K = torch.Tensor(param["intrinsics"])
if "height" not in param or "width" not in param:
assert self.cfg.image_size_during_calibration is not None
H, W = self.cfg.image_size_during_calibration
else:
H, W = param["height"], param["width"]
self.camera_ids = list(param["world_2_cam"].keys())
w2c = torch.tensor([param["world_2_cam"][k] for k in self.camera_ids]) # (N, 4, 4)
R = w2c[..., :3, :3]
T = w2c[..., :3, 3]
orientation = R.transpose(-1, -2) # (N, 3, 3)
location = R.transpose(-1, -2) @ -T[..., None] # (N, 3, 1)
# adjust how cameras distribute in the space with a global rotation
if self.cfg.align_cameras_to_axes:
orientation, location = camera.align_cameras_to_axes(
orientation, location, target_convention="opengl"
)
# modify the local orientation of cameras to fit in different camera conventions
if self.cfg.camera_convention_conversion is not None:
orientation, K = camera.convert_camera_convention(
self.cfg.camera_convention_conversion, orientation, K, H, W
)
c2w = torch.cat([orientation, location], dim=-1) # camera-to-world transformation
if self.cfg.target_extrinsic_type == "w2c":
R = orientation.transpose(-1, -2)
T = orientation.transpose(-1, -2) @ -location
w2c = torch.cat([R, T], dim=-1) # world-to-camera transformation
extrinsic = w2c
elif self.cfg.target_extrinsic_type == "c2w":
extrinsic = c2w
else:
raise NotImplementedError(f"Unknown extrinsic type: {self.cfg.target_extrinsic_type}")
self.camera_params = {}
for i, camera_id in enumerate(self.camera_ids):
self.camera_params[camera_id] = {"intrinsic": K, "extrinsic": extrinsic[i]}
def filter_division(self, division):
if division is not None:
cam_for_train = [8, 7, 9, 4, 10, 5, 13, 2, 12, 1, 14, 0]
if division == "train":
self.camera_ids = [
self.camera_ids[i]
for i in range(len(self.camera_ids))
if i in cam_for_train
]
elif division == "val":
self.camera_ids = [
self.camera_ids[i]
for i in range(len(self.camera_ids))
if i not in cam_for_train
]
elif division == "front-view":
self.camera_ids = self.camera_ids[8:9]
elif division == "side-view":
self.camera_ids = self.camera_ids[0:1]
elif division == "six-view":
self.camera_ids = [self.camera_ids[i] for i in [0, 1, 7, 8, 14, 15]]
else:
raise NotImplementedError(f"Unknown division type: {division}")
logger.info(f"division: {division}")
def apply_transforms(self, item):
if self.cfg.use_color_correction:
color_correction_path = self.cfg.root_folder / 'color_correction' / self.cfg.subject / f'{item["camera_id"]}.npy'
affine_color_transform = np.load(color_correction_path)
rgb = item["rgb"] / 255
rgb = rgb @ affine_color_transform[:3, :3] + affine_color_transform[np.newaxis, :3, 3]
item["rgb"] = (np.clip(rgb, 0, 1) * 255).astype(np.uint8)
super().apply_transforms(item)
return item
if __name__ == "__main__":
import tyro
from tqdm import tqdm
from torch.utils.data import DataLoader
from vhap.config.nersemble import NersembleDataConfig
from vhap.config.base import import_module
cfg = tyro.cli(NersembleDataConfig)
cfg.use_landmark = False
dataset = import_module(cfg._target)(
cfg=cfg,
img_to_tensor=False,
batchify_all_views=True,
)
print(len(dataset))
sample = dataset[0]
print(sample.keys())
print(sample["rgb"].shape)
dataloader = DataLoader(dataset, batch_size=None, shuffle=False, num_workers=1)
for item in tqdm(dataloader):
pass
|