isegm/data/datasets/pascalvoc.py

import pickle as pkl
from pathlib import Path

import cv2
import numpy as np

from isegm.data.base import ISDataset
from isegm.data.sample import DSample


class PascalVocDataset(ISDataset):
    def __init__(self, dataset_path, split='train', **kwargs):
        super().__init__(**kwargs)
        assert split in {'train', 'val', 'trainval', 'test'}

        self.dataset_path = Path(dataset_path)
        self._images_path = self.dataset_path / "JPEGImages"
        self._insts_path = self.dataset_path / "SegmentationObject"
        self.dataset_split = split

        if split == 'test':
            with open(self.dataset_path / f'ImageSets/Segmentation/test.pickle', 'rb') as f:
                self.dataset_samples, self.instance_ids = pkl.load(f)
        else:
            with open(self.dataset_path / f'ImageSets/Segmentation/{split}.txt', 'r') as f:
                self.dataset_samples = [name.strip() for name in f.readlines()]

    def get_sample(self, index) -> DSample:
        sample_id = self.dataset_samples[index]
        image_path = str(self._images_path / f'{sample_id}.jpg')
        mask_path = str(self._insts_path / f'{sample_id}.png')

        image = cv2.imread(image_path)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        instances_mask = cv2.imread(mask_path)
        instances_mask = cv2.cvtColor(instances_mask, cv2.COLOR_BGR2GRAY).astype(np.int32)
        if self.dataset_split == 'test':
            instance_id = self.instance_ids[index]
            mask = np.zeros_like(instances_mask)
            mask[instances_mask == 220] = 220  # ignored area
            mask[instances_mask == instance_id] = 1
            objects_ids = [1]
            instances_mask = mask
        else:
            objects_ids = np.unique(instances_mask)
            objects_ids = [x for x in objects_ids if x != 0 and x != 220]

        return DSample(image, instances_mask, objects_ids=objects_ids, ignore_ids=[220], sample_id=index)