baselines/UIED-3.3/cnn_pytorch/datasets.py

import os
import json
from PIL import Image
from torch.utils.data import Dataset
from torchvision.transforms import ToTensor


TRAIN_ANN_PATH = 'annotations/instances_train2017.json'
VAL_ANN_PATH = 'annotations/instances_val2017.json'
TEST_ANN_PATH = 'annotations/instances_test2017.json'

TRAIN_IMG_DIR = 'images/instances_train2017'
VAL_IMG_DIR = 'images/instances_val2017'
TEST_IMG_DIR = 'images/instances_test2017'


class TrainDataset(Dataset):

    def __init__(self, dataset_root, is_test = False, transform=None):
        super(Dataset, self).__init__()
        self.dataset_root = dataset_root
        img_dir = TEST_IMG_DIR if is_test else TRAIN_IMG_DIR
        ann_path = TEST_ANN_PATH if is_test else TRAIN_ANN_PATH
        self.img_dir = os.path.join(dataset_root, img_dir)
        ann_path = os.path.join(dataset_root, ann_path)
        with open(ann_path, 'r') as f:
            self.ann_data = json.load(f)
        self.transform = transform
        self.id_file_dict = {img['id']: img['file_name'] for img in self.ann_data['images']}
        
    
    def __len__(self):
        return len(self.ann_data['annotations'])
    

    def __getitem__(self, idx):
        ann = self.ann_data['annotations'][idx]
        bbox = ann['bbox']
        img = Image.open(os.path.join(self.img_dir, self.id_file_dict[ann['image_id']]))
        img = img.crop((bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]))

        if self.transform:
            img = self.transform(img)
        else:
            img = ToTensor()(img)

        return img, ann['category_id'] - 1
    

class PredDataset(Dataset):

    def __init__(self, dataset_root, region_path, transform=None):
        super(Dataset, self).__init__()
        self.dataset_root = dataset_root
        self.img_dir = os.path.join(dataset_root, TEST_IMG_DIR)
        ann_path = os.path.join(dataset_root, TEST_ANN_PATH)
        with open(ann_path, 'r') as f:
            self.ann_data = json.load(f)
        with open(region_path, 'r') as f:
            self.regions = json.load(f)
        self.transform = transform
        self.id_file_dict = {ann['id']: ann['file_name'] for ann in self.ann_data['images']}
        
    
    def __len__(self):
        return len(self.regions)
    

    def __getitem__(self, idx):
        region = self.regions[idx]
        bbox = region['bbox']
        img = Image.open(os.path.join(self.img_dir, self.id_file_dict[region['image_id']]))
        img = img.crop((bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]))

        if self.transform:
            img = self.transform(img)
        else:
            img = ToTensor()(img)

        return img