update

README.md

@@ -121,6 +121,35 @@ x: dict(
 """
 ```
 
+### LEVIR Change Detection + (LEVIR-CD+)
+
+<img src="./assets/levircd_plus.png" width="750px"></img>
+
+The [LEVIR-CD+](https://github.com/AnonymousForACMMM/Dataset) dataset, proposed in ["S2Looking: A Satellite Side-Looking Dataset for Building Change Detection", Shen et al.](https://arxiv.org/abs/2107.09244) is an urban Change Detection dataset of 985 very high resolution (VHR) 0.5m RGB image pairs extracted from Google Earth. The dataset contains building/land use change masks from 20 different regions of Texas between 2002-2020 with a temporal difference of 5 years.
+
+The dataset can be downloaded (3.6GB) using `scripts/download_levircd_plus.sh` and instantiated below:
+
+```python
+from torchrs.transforms import Compose, ToTensor
+from torchrs.datasets import LEVIRCD_Plus
+
+transform = Compose([ToTensor()])
+
+dataset = LEVIRCD_Plus(
+    root="path/to/dataset/",
+    split="train",  # or 'test'
+    transform=transform,
+)
+
+x = dataset[0]
+"""
+x: dict(
+    x: (2, 3, 1024, 1024)
+    mask: (1, 1024, 1024)
+)
+"""
+```
+
 ### Remote Sensing Visual Question Answering (RSVQA) Low Resolution (LR)
 
 <img src="./assets/rsvqa_lr.png" width="850px"></img>

torchrs/datasets/__init__.py

@@ -5,8 +5,10 @@
 from .resisc45 import RESISC45
 from .rsicd import RSICD
 from .oscd import OSCD
+from .levircd import LEVIRCD_Plus
+
 
 __all__ = [
     "PROBAV", "ETCI2021", "RSVQALR", "EuroSATRGB", "EuroSATMS",
-    "RESISC45", "RSICD", "OSCD"
+    "RESISC45", "RSICD", "OSCD", "LEVIRCD_Plus"
 ]

torchrs/datasets/levircd.py

@@ -0,0 +1,62 @@
+import os
+from glob import glob
+from typing import Dict
+
+import torch
+import numpy as np
+from PIL import Image
+
+from torchrs.transforms import Compose, ToTensor
+
+
+class LEVIRCD_Plus(torch.utils.data.Dataset):
+    """ LEVIR-CD+ dataset from 'S2Looking: A Satellite Side-Looking
+    Dataset for Building Change Detection', Shen at al. (2021)
+    https://arxiv.org/abs/2107.09244
+
+    'LEVIR-CD+ contains more than 985 VHR (0.5m/pixel) bitemporal Google
+    Earth images with dimensions of 1024x1024 pixels. These bitemporal images
+    are from 20 different regions located in several cities in the state of
+    Texas in the USA. The capture times of the image data vary from 2002 to
+    2020. Images of different regions were taken at different times. The
+    bitemporal images have a time span of 5 years.'
+    """
+    def __init__(
+        self,
+        root: str = ".data/levircd_plus",
+        split: str = "train",
+        transform: Compose = Compose([ToTensor()]),
+    ):
+        assert split in ["train", "test"]
+        self.root = root
+        self.transform = transform
+        self.files = self.load_files(root, split)
+
+    @staticmethod
+    def load_files(root: str, split: str):
+        files = []
+        images = glob(os.path.join(root, split, "A", "*.png"))
+        images = sorted([os.path.basename(image) for image in images])
+        for image in images:
+            image1 = os.path.join(root, split, "A", image)
+            image2 = os.path.join(root, split, "B", image)
+            mask = os.path.join(root, split, "label", image)
+            files.append(dict(image1=image1, image2=image2, mask=mask))
+        return files
+
+    def __len__(self) -> int:
+        return len(self.files)
+
+    def __getitem__(self, idx: int) -> Dict:
+        """ Returns a dict containing x, mask, region, dates
+        x: (2, 13, h, w)
+        mask: (1, h, w)
+        """
+        files = self.files[idx]
+        mask = np.array(Image.open(files["mask"]))
+        mask = np.clip(mask, 0, 1)
+        image1 = np.array(Image.open(files["image1"]))
+        image2 = np.array(Image.open(files["image2"]))
+        image1, image2, mask = self.transform([image1, image2, mask])
+        x = torch.stack([image1, image2], dim=0)
+        return dict(x=x, mask=mask)