File size: 9,446 Bytes
a93afca
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import random

import cv2
import numpy as np
import torch
import torch.nn.functional as F
from pycocotools import mask

from model.segment_anything.utils.transforms import ResizeLongestSide

from .grefer import G_REFER
from .refer import REFER
from torchvision import transforms


class ReferSegDataset(torch.utils.data.Dataset):
    pixel_mean = torch.Tensor([123.675, 116.28, 103.53]).view(-1, 1, 1)
    pixel_std = torch.Tensor([58.395, 57.12, 57.375]).view(-1, 1, 1)
    img_size = 1024
    ignore_label = 255

    def __init__(
        self,
        base_image_dir,
        tokenizer,
        samples_per_epoch=500 * 8 * 2 * 10,
        precision: str = "fp32",
        image_size: int = 224,
        num_classes_per_sample: int = 3,
        exclude_val=False,
        refer_seg_data="refclef||refcoco||refcoco+||refcocog",
        model_type="ori",
        transform=ResizeLongestSide(1024),
    ):
        self.model_type = model_type
        self.exclude_val = exclude_val
        self.samples_per_epoch = samples_per_epoch
        self.num_classes_per_sample = num_classes_per_sample

        self.base_image_dir = base_image_dir
        self.tokenizer = tokenizer
        self.precision = precision
        self.transform = transform
        self.image_preprocessor = transforms.Compose([
            transforms.ToTensor(),
            transforms.Resize((image_size, image_size), interpolation=3), 
            transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
        ])
        
        DATA_DIR = os.path.join(base_image_dir, "refer_seg")
        self.refer_seg_ds_list = refer_seg_data.split(
            "||"
        )  # ['refclef', 'refcoco', 'refcoco+', 'refcocog']
        self.refer_seg_data = {}
        for ds in self.refer_seg_ds_list:
            if ds == "refcocog":
                splitBy = "umd"
            else:
                splitBy = "unc"

            if ds == "grefcoco":
                refer_api = G_REFER(DATA_DIR, ds, splitBy)
            else:
                refer_api = REFER(DATA_DIR, ds, splitBy)

            ref_ids_train = refer_api.getRefIds(split="train")
            images_ids_train = refer_api.getImgIds(ref_ids=ref_ids_train)
            refs_train = refer_api.loadRefs(ref_ids=ref_ids_train)

            refer_seg_ds = {}
            refer_seg_ds["images"] = []
            loaded_images = refer_api.loadImgs(image_ids=images_ids_train)

            for item in loaded_images:
                item = item.copy()
                if ds == "refclef":
                    item["file_name"] = os.path.join(
                        DATA_DIR, "images/saiapr_tc-12", item["file_name"]
                    )
                else:
                    item["file_name"] = os.path.join(
                        DATA_DIR, "images/mscoco/images/train2014", item["file_name"]
                    )
                refer_seg_ds["images"].append(item)
            refer_seg_ds["annotations"] = refer_api.Anns  # anns_train

            print(
                "dataset {} (refs {}) (train split) has {} images and {} annotations.".format(
                    ds,
                    splitBy,
                    len(refer_seg_ds["images"]),
                    len(refer_seg_ds["annotations"]),
                )
            )

            img2refs = {}
            for ref in refs_train:
                image_id = ref["image_id"]
                img2refs[image_id] = img2refs.get(image_id, []) + [
                    ref,
                ]
            refer_seg_ds["img2refs"] = img2refs
            self.refer_seg_data[ds] = refer_seg_ds

    def __len__(self):
        return self.samples_per_epoch

    def preprocess(self, x: torch.Tensor) -> torch.Tensor:
        """Normalize pixel values and pad to a square input."""
        if self.model_type=="hq":
            h, w = x.shape[-2:]
            padh = self.img_size - h
            padw = self.img_size - w
            x = F.pad(x, (0, padw, 0, padh), value=128)
            
        # Normalize colors
        x = (x - self.pixel_mean) / self.pixel_std

        if self.model_type=="effi":
            x = F.interpolate(x.unsqueeze(0), (self.img_size, self.img_size), mode="bilinear").squeeze(0)
        else:
            # Pad
            h, w = x.shape[-2:]
            padh = self.img_size - h
            padw = self.img_size - w
            x = F.pad(x, (0, padw, 0, padh))
        return x

    def __getitem__(self, idx):
        ds = random.randint(0, len(self.refer_seg_ds_list) - 1)
        ds = self.refer_seg_ds_list[ds]
        refer_seg_ds = self.refer_seg_data[ds]
        images = refer_seg_ds["images"]
        annotations = refer_seg_ds["annotations"]
        img2refs = refer_seg_ds["img2refs"]
        idx = random.randint(0, len(images) - 1)
        image_info = images[idx]
        image_path = image_info["file_name"]
        image_id = image_info["id"]
        refs = img2refs[image_id]
        if len(refs) == 0:
            return self.__getitem__(0)

        sents = []
        ann_ids = []
        for ref in refs:
            for sent in ref["sentences"]:
                text = sent["sent"]
                sents.append(text)
                ann_ids.append(ref["ann_id"])
        if len(sents) >= self.num_classes_per_sample:
            sampled_inds = np.random.choice(
                list(range(len(sents))), size=self.num_classes_per_sample, replace=False
            )
        else:
            sampled_inds = list(range(len(sents)))
        sampled_sents = np.vectorize(sents.__getitem__)(sampled_inds).tolist()
        # sampled_ann_ids = np.vectorize(ann_ids.__getitem__)(sampled_inds).tolist()
        sampled_ann_ids = [ann_ids[ind] for ind in sampled_inds]
        sampled_classes = sampled_sents
        image = cv2.imread(image_path)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

        # preprocess image for evf
        image_evf = self.image_preprocessor(image)

        image = self.transform.apply_image(image)  # preprocess image for sam
        resize = image.shape[:2]

        image = self.preprocess(torch.from_numpy(image).permute(2, 0, 1).contiguous())

        flag = False
        masks = []
        for ann_id in sampled_ann_ids:
            if isinstance(ann_id, list):
                flag = True
                if -1 in ann_id:
                    assert len(ann_id) == 1
                    m = np.zeros((image_info["height"], image_info["width"])).astype(
                        np.uint8
                    )
                else:
                    m_final = np.zeros(
                        (image_info["height"], image_info["width"])
                    ).astype(np.uint8)
                    for ann_id_i in ann_id:
                        ann = annotations[ann_id_i]

                        if len(ann["segmentation"]) == 0:
                            m = np.zeros(
                                (image_info["height"], image_info["width"])
                            ).astype(np.uint8)
                        else:
                            if type(ann["segmentation"][0]) == list:  # polygon
                                rle = mask.frPyObjects(
                                    ann["segmentation"],
                                    image_info["height"],
                                    image_info["width"],
                                )
                            else:
                                rle = ann["segmentation"]
                                for i in range(len(rle)):
                                    if not isinstance(rle[i]["counts"], bytes):
                                        rle[i]["counts"] = rle[i]["counts"].encode()
                            m = mask.decode(rle)
                            m = np.sum(
                                m, axis=2
                            )  # sometimes there are multiple binary map (corresponding to multiple segs)
                            m = m.astype(np.uint8)  # convert to np.uint8
                        m_final = m_final | m
                    m = m_final
                masks.append(m)
                continue

            ann = annotations[ann_id]

            if len(ann["segmentation"]) == 0:
                m = np.zeros((image_info["height"], image_info["width"])).astype(
                    np.uint8
                )
                masks.append(m)
                continue

            if type(ann["segmentation"][0]) == list:  # polygon
                rle = mask.frPyObjects(
                    ann["segmentation"], image_info["height"], image_info["width"]
                )
            else:
                rle = ann["segmentation"]
                for i in range(len(rle)):
                    if not isinstance(rle[i]["counts"], bytes):
                        rle[i]["counts"] = rle[i]["counts"].encode()
            m = mask.decode(rle)
            m = np.sum(
                m, axis=2
            )  # sometimes there are multiple binary map (corresponding to multiple segs)
            m = m.astype(np.uint8)  # convert to np.uint8
            masks.append(m)

        masks = np.stack(masks, axis=0)

        masks = torch.from_numpy(masks)
        label = torch.ones(masks.shape[1], masks.shape[2]) * self.ignore_label

        return (
            image_path,
            image,
            image_evf,
            masks,
            label,
            resize,
            sampled_classes,
        )