ObjectSizeEstimation/dataset.py at main · Amandahsr/ObjectSizeEstimation · GitHub

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
from typing import Dict, List, Tuple

import tensorflow as tf
import tensorflow_datasets as tfds

from utils import ImageEnhancements


class KittiDataset:
    def __init__(self, image_enhancer: ImageEnhancements, image_enhance: bool = False):
        self.input_shape: Tuple[int, int, int] = (224, 224, 3)
        self.datatype: tf.DType = tf.float32
        self.batch_size: int = 64
        self.shuffle_seed: int = 15
        self.image_enhance = image_enhance

        self.image_enhancer = image_enhancer
        self.training_dataset: List[Dict] = None
        self.validation_dataset: List[Dict] = None
        self.testing_dataset: List[Dict] = None

        self.load_dataset()
        self.process_datasets()

    def load_dataset(self) -> None:
        """
        Loads Kitti training and testing datasets.
        """
        self.training_dataset = tfds.load("kitti", split="train", as_supervised=False)
        self.validation_dataset = tfds.load(
            "kitti", split="validation", as_supervised=False
        )
        self.testing_dataset = tfds.load("kitti", split="test", as_supervised=False)

    def _enhance_batch_dataset(self, x, y) -> Dict:
        """
        Applies image enhancements to a batch of images using ImageEnhancements class.
        """
        enhanced_images = tf.map_fn(
            self.image_enhancer.image_enhance_tensor,
            x["image"],
            fn_output_signature=tf.float32,
        )

        enhanced_images.set_shape(x["image"].shape)

        enhanced_x = {
            "image": enhanced_images,
            "box_metadata": x["box_metadata"],
        }

        return enhanced_x, y

    def _crop_image_to_single_obj(self, data: Dict) -> tf.data.Dataset:
        """
        Crops multi-object images to single objects using bounding box annotations.
        """
        # Extract annotations from data
        image = tf.cast(data["image"], self.datatype) / 255
        boxes = tf.cast(data["objects"]["bbox"], self.datatype)
        dimensions_3d = tf.cast(data["objects"]["dimensions"], self.datatype)

        # Calculate xy location of bounding boxes in image, helps with actual geometry estimations
        ymin = boxes[:, 0]
        xmin = boxes[:, 1]
        ymax = boxes[:, 2]
        xmax = boxes[:, 3]

        box_width = xmax - xmin
        box_height = ymax - ymin
        box_center_x = (xmin + xmax) / 2
        box_center_y = (ymin + ymax) / 2

        box_xy_metadata = tf.stack(
            [
                box_width,
                box_height,
                box_center_x,
                box_center_y,
            ],
            axis=1,
        )

        # Crop image to single 2d object images
        num_objects_in_img = tf.shape(boxes)[0]
        image_batch = tf.expand_dims(image, axis=0)
        box_indices = tf.zeros((num_objects_in_img,), dtype=tf.int32)

        crops = tf.image.crop_and_resize(
            image=image_batch,
            boxes=boxes,
            box_indices=box_indices,
            crop_size=self.input_shape[:2],
            method="bilinear",
        )

        return tf.data.Dataset.from_tensor_slices(
            (
                {
                    "image": crops,
                    "box_metadata": box_xy_metadata,
                },
                dimensions_3d,
            )
        )

    def _prepare_dataset(self, dataset, training: bool):
        """
        Convert images to single-object image dataset.
        """
        dataset = dataset.flat_map(self._crop_image_to_single_obj)

        if training:
            dataset = dataset.shuffle(
                buffer_size=4096,
                seed=self.shuffle_seed,
                reshuffle_each_iteration=True,
            )

        dataset = dataset.batch(self.batch_size)

        if self.image_enhance:
            dataset = dataset.map(
                self._enhance_batch_dataset,
                num_parallel_calls=tf.data.AUTOTUNE,
            )

        dataset = dataset.prefetch(tf.data.AUTOTUNE)

        return dataset

    def process_datasets(self) -> None:
        """
        Processes training, validation and testing datasets for object size estimations.
        """
        self.training_dataset = self._prepare_dataset(
            self.training_dataset,
            training=True,
        )

        self.validation_dataset = self._prepare_dataset(
            self.validation_dataset,
            training=False,
        )

        self.testing_dataset = self._prepare_dataset(
            self.testing_dataset,
            training=False,
        )