-
Notifications
You must be signed in to change notification settings - Fork 42
/
flame_dataset.py
211 lines (183 loc) · 8.68 KB
/
flame_dataset.py
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
import os
import json
from typing import Dict, Any, List, Union, Tuple, Optional
from collections import namedtuple
import torch
from torch.utils.data import Dataset
import numpy as np
from hydra.utils import instantiate
import albumentations as A
import pytorch_toolbelt.utils as pt_utils
from model_training.data.config import (
IMAGE_FILENAME_KEY,
SAMPLE_INDEX_KEY,
INPUT_IMAGE_KEY,
INPUT_BBOX_KEY,
INPUT_SIZE_KEY,
TARGET_PROJECTION_MATRIX,
TARGET_3D_MODEL_VERTICES,
TARGET_3D_WORLD_VERTICES,
TARGET_2D_LANDMARKS,
TARGET_LANDMARKS_HEATMAP,
TARGET_2D_FULL_LANDMARKS,
TARGET_2D_LANDMARKS_PRESENCE,
)
from model_training.data.transforms import get_resize_fn, get_normalize_fn
from model_training.data.utils import ensure_bbox_boundaries, extend_bbox, read_as_rgb, get_68_landmarks
from model_training.utils import load_2d_indices, create_logger
MeshArrays = namedtuple(
"MeshArrays",
["vertices3d", "vertices3d_world_homo", "projection_matrix"],
)
logger = create_logger(__name__)
def collate_skip_none(batch: Any) -> Any:
len_batch = len(batch)
batch = list(filter(lambda x: x is not None, batch))
if len_batch > len(batch):
diff = len_batch - len(batch)
batch = batch + batch[:diff]
return torch.utils.data.dataloader.default_collate(batch)
class FlameDataset(Dataset):
def __init__(self, data: List[Dict[str, Any]], config: Dict[str, Any]) -> None:
self.data = data
self.config = config
self.img_size = config["img_size"]
self.filename_key = "img_path"
self.aug_pipeline = self._get_aug_pipeline(config["transform"])
self.num_classes = config.get("num_classes")
self.keypoints_indices = load_2d_indices(config["keypoints"])
self.tensor_keys = [INPUT_IMAGE_KEY]
self.coder = instantiate(config["coder"], config, self.num_classes)
def __len__(self) -> int:
return len(self.data)
def __getitem__(self, idx: int) -> Optional[Dict[str, Any]]:
item_anno = self._get_item_anno(idx=idx)
item_data = self._parse_anno(item_anno)
item_data = self._transform(item_data)
item_dict = self._form_anno_dict(item_data)
item_dict = self._add_index(idx, item_anno, item_dict)
item_dict = self._convert_images_to_tensors(item_dict)
return item_dict
def _add_index(self, idx: int, annotation: Any, item_dict: Dict[str, Any]) -> Dict[str, Any]:
if item_dict is not None:
item_dict.update({SAMPLE_INDEX_KEY: idx, IMAGE_FILENAME_KEY: annotation[self.filename_key]})
return item_dict
def _get_item_anno(self, idx: int) -> Dict[str, Any]:
return self.data[idx]
@classmethod
def from_config(cls, config: Dict[str, Any]):
with open(config["ann_path"]) as json_file:
anno = json.load(json_file)
return cls(data=anno, config=config)
def _convert_images_to_tensors(self, item_data: Dict[str, Any]) -> Dict[str, Any]:
if item_data is not None:
for key, item in item_data.items():
if isinstance(item, np.ndarray) and key in self.tensor_keys:
item_data[key] = pt_utils.image_to_tensor(item.astype("float32"))
return item_data
def _parse_anno(self, item_anno: Dict[str, Any]) -> Dict[str, Any]:
img = read_as_rgb(os.path.join(self.config["dataset_root"], item_anno["img_path"]))
bbox = item_anno["bbox"]
offset = tuple(0.1 * np.random.uniform(size=4) + 0.05)
x, y, w, h = ensure_bbox_boundaries(extend_bbox(np.array(bbox), offset), img.shape[:2])
cropped_img = img[y : y + h, x : x + w]
(
flame_vertices3d,
flame_vertices3d_world_homo,
projection_matrix,
) = self._load_mesh(os.path.join(self.config["dataset_root"], item_anno["annotation_path"]))
return {
INPUT_IMAGE_KEY: cropped_img,
INPUT_BBOX_KEY: (x, y, w, h),
INPUT_SIZE_KEY: img.shape,
TARGET_3D_MODEL_VERTICES: flame_vertices3d,
TARGET_3D_WORLD_VERTICES: flame_vertices3d_world_homo,
TARGET_PROJECTION_MATRIX: projection_matrix
}
@staticmethod
def _load_mesh(mesh_path: str) -> MeshArrays:
with open(mesh_path) as json_data:
data = json.load(json_data)
flame_vertices3d = np.array(data["vertices"], dtype=np.float32)
model_view_matrix = np.array(data["model_view_matrix"], dtype=np.float32)
flame_vertices3d_homo = np.concatenate((flame_vertices3d, np.ones_like(flame_vertices3d[:, [0]])), -1)
# rotated and translated (to world coordinates)
flame_vertices3d_world_homo = np.transpose(np.matmul(model_view_matrix, np.transpose(flame_vertices3d_homo)))
return MeshArrays(
vertices3d=flame_vertices3d,
vertices3d_world_homo=flame_vertices3d_world_homo, # with pose and translation
projection_matrix=np.array(data["projection_matrix"], dtype=np.float32),
)
@staticmethod
def _project_vertices_onto_image(
vertices3d_world_homo: np.ndarray,
projection_matrix: np.ndarray,
height: int,
crop_point_x: int,
crop_point_y: int
):
vertices2d_homo = np.transpose(np.matmul(projection_matrix, np.transpose(vertices3d_world_homo)))
vertices2d = vertices2d_homo[:, :2] / vertices2d_homo[:, [3]]
vertices2d = np.stack((vertices2d[:, 0], (height - vertices2d[:, 1])), -1)
vertices2d -= (crop_point_x, crop_point_y)
return vertices2d
def _get_2d_landmarks_w_presence(
self,
vertices3d_world_homo: np.ndarray,
projection_matrix: np.ndarray,
img_shape: np.ndarray,
bbox: Tuple[int, int, int, int],
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
if self.num_classes == 68:
landmarks_3d_world_subset = get_68_landmarks(
torch.from_numpy(vertices3d_world_homo[..., :3]).view(-1, 3)
).numpy()
landmarks_3d_world_subset = np.concatenate(
(landmarks_3d_world_subset, np.ones_like(landmarks_3d_world_subset[:, [0]])), -1
)
else:
landmarks_3d_world_subset = vertices3d_world_homo[self.keypoints_indices]
x, y, w, h = bbox
landmarks_2d_subset = self._project_vertices_onto_image(
landmarks_3d_world_subset, projection_matrix, img_shape[0], x, y
)
keypoints_2d = self._project_vertices_onto_image(vertices3d_world_homo, projection_matrix, img_shape[0], x, y)
presence_subset = np.array([False] * len(landmarks_2d_subset))
for i in range(len(landmarks_2d_subset)):
if 0 < landmarks_2d_subset[i, 0] < w and 0 < landmarks_2d_subset[i, 1] < h:
presence_subset[i] = True
return landmarks_2d_subset, presence_subset, keypoints_2d
def _transform(self, item_data: Dict[str, Any]) -> Dict[str, Any]:
vertices_2d_subset, presence_subset, vertices_2d = self._get_2d_landmarks_w_presence(
item_data[TARGET_3D_WORLD_VERTICES],
item_data[TARGET_PROJECTION_MATRIX],
item_data[INPUT_SIZE_KEY],
item_data[INPUT_BBOX_KEY],
)
result = self.aug_pipeline(
image=item_data[INPUT_IMAGE_KEY], keypoints=np.concatenate((vertices_2d_subset, vertices_2d), 0)
)
return {
INPUT_IMAGE_KEY: result["image"],
INPUT_BBOX_KEY: item_data[INPUT_BBOX_KEY],
TARGET_3D_MODEL_VERTICES: item_data[TARGET_3D_MODEL_VERTICES],
TARGET_2D_LANDMARKS: np.array(result["keypoints"][: self.num_classes], dtype=np.float32),
TARGET_2D_FULL_LANDMARKS: np.array(result["keypoints"][self.num_classes :], dtype=np.float32),
TARGET_2D_LANDMARKS_PRESENCE: presence_subset
}
def _form_anno_dict(self, item_data: Dict[str, np.ndarray]) -> Dict[str, Union[torch.Tensor, np.ndarray]]:
landmarks = item_data[TARGET_2D_LANDMARKS]
presence = item_data[TARGET_2D_LANDMARKS_PRESENCE]
heatmap = self.coder(landmarks, presence)
item_data[TARGET_2D_LANDMARKS] = landmarks / self.img_size
item_data[TARGET_LANDMARKS_HEATMAP] = np.uint8(255.0 * heatmap)
return item_data
def _get_aug_pipeline(self, aug_config: Dict[str, Any]) -> A.Compose:
normalize = get_normalize_fn(aug_config.get("normalize", "imagenet"))
resize = get_resize_fn(self.img_size, mode=aug_config.get("resize_mode", "longest_max_size"))
return A.Compose(
[resize, normalize],
keypoint_params=A.KeypointParams(format="xy", remove_invisible=False)
)
def get_collate_fn(self) -> Any:
return collate_skip_none