Add VHR10 datamodule (#1082)

* Add VHR10 datamodule

* Add newline

* patch_size accepts int and tuple of ints

* Update conf

* VHR10 Datamodule v2

* Remove auto_lr_find

* Remove preprocess

* Update config

* Remove setting of matplotlib backend

* Remove import

* Typing update

* Key fix

* Coverage fix

* Update conf

* Update conf

* Dowload=True

* Use weights

* Empty commit

* Switch to ndim

* Remove conf, tight_layout and spacing

* Set constrained layout via rcParams

* Revert and bump min matplotlib version

* Switch back to dataset_split

* Separate out AugPipe

* Increase figsize & revert matplotlib

* Common collate_fn

* Class var std

* Undo std change in BaseDataModule

* Undo req changes

* Remove unused line

* Add version strings

* mypy fix

---------

Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>

tests/conf/vhr10.yaml

@@ -0,0 +1,17 @@
+model:
+  class_path: ObjectDetectionTask
+  init_args:
+    model: "faster-rcnn"
+    backbone: "resnet50"
+    num_classes: 11
+    lr: 2.5e-5
+    patience: 10
+data:
+  class_path: VHR10DataModule
+  init_args:
+    batch_size: 1
+    num_workers: 0
+    patch_size: 4
+  dict_kwargs:
+    root: "tests/data/vhr10"
+    download: true

tests/data/vhr10/annotations.json

@@ -1 +1 @@
-{"images": [{"file_name": "001.jpg", "height": 8, "width": 8, "id": 0}, {"file_name": "002.jpg", "height": 8, "width": 8, "id": 1}, {"file_name": "003.jpg", "height": 8, "width": 8, "id": 2}, {"file_name": "004.jpg", "height": 8, "width": 8, "id": 3}, {"file_name": "005.jpg", "height": 8, "width": 8, "id": 4}], "annotations": [{"id": 0, "image_id": 0, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "iscrowd": 0}, {"id": 1, "image_id": 1, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 2, "image_id": 2, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 3, "image_id": 3, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 4, "image_id": 4, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}]}
+{"images": [{"file_name": "001.jpg", "height": 8, "width": 8, "id": 0}, {"file_name": "002.jpg", "height": 8, "width": 8, "id": 1}, {"file_name": "003.jpg", "height": 8, "width": 8, "id": 2}, {"file_name": "004.jpg", "height": 8, "width": 8, "id": 3}, {"file_name": "005.jpg", "height": 8, "width": 8, "id": 4}], "annotations": [{"id": 0, "image_id": 0, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 1, "image_id": 1, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 2, "image_id": 2, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 3, "image_id": 3, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}, {"id": 4, "image_id": 4, "category_id": 1, "area": 4.0, "bbox": [4, 4, 2, 2], "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]], "iscrowd": 0}]}

tests/data/vhr10/data.py

@@ -5,7 +5,6 @@ import json
 import os
 import shutil
 import subprocess
-from copy import deepcopy
 
 import numpy as np
 from PIL import Image
@@ -47,7 +46,7 @@ def generate_test_data(root: str, n_imgs: int = 3) -> str:
         )
 
     ann = 0
-    for i, img in enumerate(ANNOTATION_FILE["images"]):
+    for _, img in enumerate(ANNOTATION_FILE["images"]):
         annot = {
             "id": ann,
             "image_id": img["id"],
@@ -57,12 +56,7 @@ def generate_test_data(root: str, n_imgs: int = 3) -> str:
             "segmentation": [[1, 1, 2, 2, 3, 3, 4, 5, 5]],
             "iscrowd": 0,
         }
-        if i != 0:
-            ANNOTATION_FILE["annotations"].append(annot)
-        else:
-            noseg_annot = deepcopy(annot)
-            del noseg_annot["segmentation"]
-            ANNOTATION_FILE["annotations"].append(noseg_annot)
+        ANNOTATION_FILE["annotations"].append(annot)
         ann += 1
 
     with open(ann_file, "w") as j:

tests/datasets/test_vhr10.py

@@ -35,11 +35,11 @@ class TestVHR10:
         monkeypatch.setattr(torchgeo.datasets.utils, "download_url", download_url)
         url = os.path.join("tests", "data", "vhr10", "NWPU VHR-10 dataset.rar")
         monkeypatch.setitem(VHR10.image_meta, "url", url)
-        md5 = "5fddb0dfd56a80638831df9f90cbf37a"
+        md5 = "92769845cae6a4e8c74bfa1a0d1d4a80"
         monkeypatch.setitem(VHR10.image_meta, "md5", md5)
         url = os.path.join("tests", "data", "vhr10", "annotations.json")
         monkeypatch.setitem(VHR10.target_meta, "url", url)
-        md5 = "833899cce369168e0d4ee420dac326dc"
+        md5 = "567c4cd8c12624864ff04865de504c58"
         monkeypatch.setitem(VHR10.target_meta, "md5", md5)
         root = str(tmp_path)
         split = request.param

tests/trainers/test_detection.py

@@ -67,7 +67,7 @@ def plot(*args: Any, **kwargs: Any) -> None:
 
 
 class TestObjectDetectionTask:
-    @pytest.mark.parametrize("name", ["nasa_marine_debris"])
+    @pytest.mark.parametrize("name", ["nasa_marine_debris", "vhr10"])
     @pytest.mark.parametrize("model_name", ["faster-rcnn", "fcos", "retinanet"])
     def test_trainer(
         self, monkeypatch: MonkeyPatch, name: str, model_name: str, fast_dev_run: bool

tests/transforms/test_transforms.py

@@ -23,7 +23,7 @@ def batch_gray() -> dict[str, Tensor]:
     return {
         "image": torch.tensor([[[[1, 2, 3], [4, 5, 6], [7, 8, 9]]]], dtype=torch.float),
         "mask": torch.tensor([[[0, 0, 1], [0, 1, 1], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[0, 1], [1, 1], [1, 0], [0, 0]]], dtype=torch.float),
+        "boxes": torch.tensor([[0.0, 0.0, 2.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
 
@@ -42,7 +42,7 @@ def batch_rgb() -> dict[str, Tensor]:
             dtype=torch.float,
         ),
         "mask": torch.tensor([[[0, 0, 1], [0, 1, 1], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[0, 1], [1, 1], [1, 0], [0, 0]]], dtype=torch.float),
+        "boxes": torch.tensor([[0.0, 0.0, 2.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
 
@@ -63,7 +63,7 @@ def batch_multispectral() -> dict[str, Tensor]:
             dtype=torch.float,
         ),
         "mask": torch.tensor([[[0, 0, 1], [0, 1, 1], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[0, 1], [1, 1], [1, 0], [0, 0]]], dtype=torch.float),
+        "boxes": torch.tensor([[0.0, 0.0, 2.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
 
@@ -79,7 +79,7 @@ def test_augmentation_sequential_gray(batch_gray: dict[str, Tensor]) -> None:
     expected = {
         "image": torch.tensor([[[[3, 2, 1], [6, 5, 4], [9, 8, 7]]]], dtype=torch.float),
         "mask": torch.tensor([[[1, 0, 0], [1, 1, 0], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[1, 0], [2, 0], [2, 1], [1, 1]]], dtype=torch.float),
+        "boxes": torch.tensor([[1.0, 0.0, 3.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
     augs = transforms.AugmentationSequential(
@@ -102,7 +102,7 @@ def test_augmentation_sequential_rgb(batch_rgb: dict[str, Tensor]) -> None:
             dtype=torch.float,
         ),
         "mask": torch.tensor([[[1, 0, 0], [1, 1, 0], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[1, 0], [2, 0], [2, 1], [1, 1]]], dtype=torch.float),
+        "boxes": torch.tensor([[1.0, 0.0, 3.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
     augs = transforms.AugmentationSequential(
@@ -129,7 +129,7 @@ def test_augmentation_sequential_multispectral(
             dtype=torch.float,
         ),
         "mask": torch.tensor([[[1, 0, 0], [1, 1, 0], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[1, 0], [2, 0], [2, 1], [1, 1]]], dtype=torch.float),
+        "boxes": torch.tensor([[1.0, 0.0, 3.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
     augs = transforms.AugmentationSequential(
@@ -156,7 +156,7 @@ def test_augmentation_sequential_image_only(
             dtype=torch.float,
         ),
         "mask": torch.tensor([[[0, 0, 1], [0, 1, 1], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[0, 1], [1, 1], [1, 0], [0, 0]]], dtype=torch.float),
+        "boxes": torch.tensor([[0.0, 0.0, 2.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
     augs = transforms.AugmentationSequential(
@@ -188,7 +188,7 @@ def test_sequential_transforms_augmentations(
             dtype=torch.float,
         ),
         "mask": torch.tensor([[[1, 0, 0], [1, 1, 0], [1, 1, 1]]], dtype=torch.long),
-        "boxes": torch.tensor([[[1, 0], [2, 0], [2, 1], [1, 1]]], dtype=torch.float),
+        "boxes": torch.tensor([[1.0, 0.0, 3.0, 2.0]], dtype=torch.float),
         "labels": torch.tensor([[0, 1]]),
     }
     train_transforms = transforms.AugmentationSequential(

torchgeo/datamodules/__init__.py

@@ -38,6 +38,7 @@ from .ucmerced import UCMercedDataModule
 from .usavars import USAVarsDataModule
 from .utils import MisconfigurationException
 from .vaihingen import Vaihingen2DDataModule
+from .vhr10 import VHR10DataModule
 from .xview import XView2DataModule
 
 __all__ = (
@@ -79,6 +80,7 @@ __all__ = (
     "UCMercedDataModule",
     "USAVarsDataModule",
     "Vaihingen2DDataModule",
+    "VHR10DataModule",
     "XView2DataModule",
     # Base classes
     "BaseDataModule",

torchgeo/datamodules/nasa_marine_debris.py

@@ -5,28 +5,13 @@
 
 from typing import Any
 
+import kornia.augmentation as K
 import torch
-from torch import Tensor
 
 from ..datasets import NASAMarineDebris
+from ..transforms import AugmentationSequential
 from .geo import NonGeoDataModule
-from .utils import dataset_split
-
-
-def collate_fn(batch: list[dict[str, Tensor]]) -> dict[str, Any]:
-    """Custom object detection collate fn to handle variable boxes.
-
-    Args:
-        batch: list of sample dicts return by dataset
-
-    Returns:
-        batch dict output
-    """
-    output: dict[str, Any] = {}
-    output["image"] = torch.stack([sample["image"] for sample in batch])
-    output["boxes"] = [sample["boxes"] for sample in batch]
-    output["labels"] = [torch.tensor([1] * len(sample["boxes"])) for sample in batch]
-    return output
+from .utils import AugPipe, collate_fn_detection, dataset_split
 
 
 class NASAMarineDebrisDataModule(NonGeoDataModule):
@@ -35,6 +20,8 @@ class NASAMarineDebrisDataModule(NonGeoDataModule):
     .. versionadded:: 0.2
     """
 
+    std = torch.tensor(255)
+
     def __init__(
         self,
         batch_size: int = 64,
@@ -58,7 +45,14 @@ class NASAMarineDebrisDataModule(NonGeoDataModule):
         self.val_split_pct = val_split_pct
         self.test_split_pct = test_split_pct
 
-        self.collate_fn = collate_fn
+        self.aug = AugPipe(
+            AugmentationSequential(
+                K.Normalize(mean=self.mean, std=self.std), data_keys=["image", "boxes"]
+            ),
+            batch_size,
+        )
+
+        self.collate_fn = collate_fn_detection
 
     def setup(self, stage: str) -> None:
         """Set up datasets.

torchgeo/datamodules/utils.py

@@ -5,10 +5,13 @@
 
 import math
 from collections.abc import Iterable
-from typing import Any, Optional, Union
+from typing import Any, Callable, Optional, Union
 
 import numpy as np
-from torch import Generator
+import torch
+from einops import rearrange
+from torch import Generator, Tensor
+from torch.nn import Module
 from torch.utils.data import Subset, TensorDataset, random_split
 
 from ..datasets import NonGeoDataset
@@ -19,6 +22,86 @@ class MisconfigurationException(Exception):
     """Exception used to inform users of misuse with Lightning."""
 
 
+class AugPipe(Module):
+    """Pipeline for applying augmentations sequentially on select data keys.
+
+    .. versionadded:: 0.6
+    """
+
+    def __init__(
+        self, augs: Callable[[dict[str, Any]], dict[str, Any]], batch_size: int
+    ) -> None:
+        """Initialize a new AugPipe instance.
+
+        Args:
+            augs: Augmentations to apply.
+            batch_size: Batch size
+        """
+        super().__init__()
+        self.augs = augs
+        self.batch_size = batch_size
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Apply the augmentation.
+
+        Args:
+            batch: Input batch.
+
+        Returns:
+            Augmented batch.
+        """
+        batch_len = len(batch["image"])
+        for bs in range(batch_len):
+            batch_dict = {
+                "image": batch["image"][bs],
+                "labels": batch["labels"][bs],
+                "boxes": batch["boxes"][bs],
+            }
+
+            if "masks" in batch:
+                batch_dict["masks"] = batch["masks"][bs]
+
+            batch_dict = self.augs(batch_dict)
+
+            batch["image"][bs] = batch_dict["image"]
+            batch["labels"][bs] = batch_dict["labels"]
+            batch["boxes"][bs] = batch_dict["boxes"]
+
+            if "masks" in batch:
+                batch["masks"][bs] = batch_dict["masks"]
+
+        # Stack images
+        batch["image"] = rearrange(batch["image"], "b () c h w -> b c h w")
+
+        return batch
+
+
+def collate_fn_detection(batch: list[dict[str, Tensor]]) -> dict[str, Any]:
+    """Custom collate fn for object detection and instance segmentation.
+
+    Args:
+        batch: list of sample dicts return by dataset
+
+    Returns:
+        batch dict output
+
+    .. versionadded:: 0.6
+    """
+    output: dict[str, Any] = {}
+    output["image"] = [sample["image"] for sample in batch]
+    output["boxes"] = [sample["boxes"].float() for sample in batch]
+    if "labels" in batch[0]:
+        output["labels"] = [sample["labels"] for sample in batch]
+    else:
+        output["labels"] = [
+            torch.tensor([1] * len(sample["boxes"])) for sample in batch
+        ]
+
+    if "masks" in batch[0]:
+        output["masks"] = [sample["masks"] for sample in batch]
+    return output
+
+
 def dataset_split(
     dataset: Union[TensorDataset, NonGeoDataset],
     val_pct: float,

torchgeo/datamodules/vhr10.py

@@ -0,0 +1,83 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+"""NWPU VHR-10 datamodule."""
+
+from typing import Any, Union
+
+import kornia.augmentation as K
+import torch
+
+from ..datasets import VHR10
+from ..samplers.utils import _to_tuple
+from ..transforms import AugmentationSequential
+from .geo import NonGeoDataModule
+from .utils import AugPipe, collate_fn_detection, dataset_split
+
+
+class VHR10DataModule(NonGeoDataModule):
+    """LightningDataModule implementation for the VHR10 dataset.
+
+    .. versionadded:: 0.6
+    """
+
+    std = torch.tensor(255)
+
+    def __init__(
+        self,
+        batch_size: int = 64,
+        patch_size: Union[tuple[int, int], int] = 512,
+        num_workers: int = 0,
+        val_split_pct: float = 0.2,
+        test_split_pct: float = 0.2,
+        **kwargs: Any,
+    ) -> None:
+        """Initialize a new VHR10DataModule instance.
+
+        Args:
+            batch_size: Size of each mini-batch.
+            patch_size: Size of each patch, either ``size`` or ``(height, width)``.
+            num_workers: Number of workers for parallel data loading.
+            val_split_pct: Percentage of the dataset to use as a validation set.
+            test_split_pct: Percentage of the dataset to use as a test set.
+            **kwargs: Additional keyword arguments passed to
+                :class:`~torchgeo.datasets.VHR10`.
+        """
+        super().__init__(VHR10, batch_size, num_workers, **kwargs)
+
+        self.val_split_pct = val_split_pct
+        self.test_split_pct = test_split_pct
+        self.patch_size = _to_tuple(patch_size)
+
+        self.collate_fn = collate_fn_detection
+
+        self.train_aug = AugPipe(
+            AugmentationSequential(
+                K.Normalize(mean=self.mean, std=self.std),
+                K.Resize(self.patch_size),
+                K.RandomHorizontalFlip(),
+                K.ColorJiggle(0.1, 0.1, 0.1, 0.1, p=0.7),
+                K.RandomVerticalFlip(),
+                data_keys=["image", "boxes", "masks"],
+            ),
+            batch_size,
+        )
+        self.aug = AugPipe(
+            AugmentationSequential(
+                K.Normalize(mean=self.mean, std=self.std),
+                K.Resize(self.patch_size),
+                data_keys=["image", "boxes", "masks"],
+            ),
+            batch_size,
+        )
+
+    def setup(self, stage: str) -> None:
+        """Set up datasets.
+
+        Args:
+            stage: Either 'fit', 'validate', 'test', or 'predict'.
+        """
+        self.dataset = VHR10(**self.kwargs)
+        self.train_dataset, self.val_dataset, self.test_dataset = dataset_split(
+            self.dataset, self.val_split_pct, self.test_split_pct
+        )

torchgeo/datasets/vhr10.py

@@ -45,10 +45,7 @@ def convert_coco_poly_to_mask(
         mask = torch.as_tensor(mask, dtype=torch.uint8)
         mask = mask.any(dim=2)
         masks.append(mask)
-    if masks:
-        masks_tensor = torch.stack(masks, dim=0)
-    else:
-        masks_tensor = torch.zeros((0, height, width), dtype=torch.uint8)
+    masks_tensor = torch.stack(masks, dim=0)
     return masks_tensor
 
 
@@ -89,10 +86,8 @@ class ConvertCocoAnnotations:
         categories = [obj["category_id"] for obj in anno]
         classes = torch.tensor(categories, dtype=torch.int64)
 
-        if "segmentation" in anno[0]:
-            segmentations = [obj["segmentation"] for obj in anno]
-        else:
-            segmentations = []
+        segmentations = [obj["segmentation"] for obj in anno]
+
         masks = convert_coco_poly_to_mask(segmentations, h, w)
 
         keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
@@ -258,8 +253,7 @@ class VHR10(NonGeoDataset):
             sample = self.coco_convert(sample)
             sample["labels"] = sample["label"]["labels"]
             sample["boxes"] = sample["label"]["boxes"]
-            if "masks" in sample["label"]:
-                sample["masks"] = sample["label"]["masks"]
+            sample["masks"] = sample["label"]["masks"]
             del sample["label"]
 
         if self.transforms is not None:
@@ -296,6 +290,7 @@ class VHR10(NonGeoDataset):
         with Image.open(filename) as img:
             array: "np.typing.NDArray[np.int_]" = np.array(img)
             tensor = torch.from_numpy(array)
+            tensor = tensor.float()
             # Convert from HxWxC to CxHxW
             tensor = tensor.permute((2, 0, 1))
             return tensor
@@ -439,7 +434,7 @@ class VHR10(NonGeoDataset):
             ncols += 1
 
         # Display image
-        fig, axs = plt.subplots(ncols=ncols, squeeze=False, figsize=(ncols * 10, 10))
+        fig, axs = plt.subplots(ncols=ncols, squeeze=False, figsize=(ncols * 10, 13))
         axs[0, 0].imshow(image)
         axs[0, 0].axis("off")
 
@@ -536,9 +531,9 @@ class VHR10(NonGeoDataset):
             if show_titles:
                 axs[0, 1].set_title("Prediction")
 
-        plt.tight_layout()
-
         if suptitle is not None:
             plt.suptitle(suptitle)
 
+        plt.tight_layout()
+
         return fig

torchgeo/transforms/transforms.py

@@ -10,6 +10,7 @@ import torch
 from einops import rearrange
 from kornia.contrib import extract_tensor_patches
 from kornia.geometry import crop_by_indices
+from kornia.geometry.boxes import Boxes
 from torch import Tensor
 from torch.nn.modules import Module
 
@@ -47,6 +48,8 @@ class AugmentationSequential(Module):
                 keys.append("input")
             elif key == "boxes":
                 keys.append("bbox")
+            elif key == "masks":
+                keys.append("mask")
             else:
                 keys.append(key)
 
@@ -67,10 +70,19 @@ class AugmentationSequential(Module):
             dtype[key] = batch[key].dtype
             batch[key] = batch[key].float()
 
+        # Convert shape of boxes from [N, 4] to [N, 4, 2]
+        if "boxes" in batch and (
+            isinstance(batch["boxes"], list) or batch["boxes"].ndim == 2
+        ):
+            batch["boxes"] = Boxes.from_tensor(batch["boxes"]).data
+
         # Kornia requires masks to have a channel dimension
-        if "mask" in batch and len(batch["mask"].shape) == 3:
+        if "mask" in batch and batch["mask"].ndim == 3:
             batch["mask"] = rearrange(batch["mask"], "b h w -> b () h w")
 
+        if "masks" in batch and batch["masks"].ndim == 3:
+            batch["masks"] = rearrange(batch["masks"], "c h w -> () c h w")
+
         inputs = [batch[k] for k in self.data_keys]
         outputs_list: Union[Tensor, list[Tensor]] = self.augs(*inputs)
         outputs_list = (
@@ -85,9 +97,17 @@ class AugmentationSequential(Module):
         for key in self.data_keys:
             batch[key] = batch[key].to(dtype[key])
 
+        # Convert boxes to default [N, 4]
+        if "boxes" in batch:
+            batch["boxes"] = Boxes(batch["boxes"]).to_tensor(
+                mode="xyxy"
+            )  # type:ignore[assignment]
+
         # Torchmetrics does not support masks with a channel dimension
         if "mask" in batch and batch["mask"].shape[1] == 1:
             batch["mask"] = rearrange(batch["mask"], "b () h w -> b h w")
+        if "masks" in batch and batch["masks"].ndim == 4:
+            batch["masks"] = rearrange(batch["masks"], "() c h w -> c h w")
 
         return batch