Merge pull request #226 from Visual-Behavior/woodscape

new: woodscape dataset

Author: Data-Iab

alodataset/__init__.py

@@ -14,3 +14,5 @@
 from .sintel_flow_dataset import SintelFlowDataset
 from .sintel_disparity_dataset import SintelDisparityDataset
 from .sintel_multi_dataset import SintelMultiDataset
+from .woodScape_dataset import WooodScapeDataset
+from .woodScape_split_dataset import WoodScapeSplitDataset

alodataset/woodScape_dataset.py

@@ -0,0 +1,253 @@
+from aloscene import Frame, Mask, BoundingBoxes2D, Labels
+from alodataset import BaseDataset
+
+from PIL import Image
+import numpy as np
+import torch
+import glob
+import os
+
+
+class WooodScapeDataset(BaseDataset):
+    """WoodScape dataset iterator
+    
+    Paramneters 
+    -----------
+        labels : List[str]
+            List of labels to stick to the frame. If the list is empty all labels are loaded. By default all labels are attached.
+        cameras : List[str]
+            List of cameras to consider. If the list empty all cameras are loaded. By default all camera views are considered.
+        fragment : Union[int, float]
+            Either the portion of dataset to to consider if the arg is float or the number of samples if int.
+            Passing a negative value will start the count from the end. By default 0.9
+        seg_classes : List[sstr]
+            Classes to consider for segmentation. By default all classes are considered.
+        merge_classees : bool
+            Assign the same classe index for all segementation classes, Default if False.
+        rename_merged : str
+            Name to give to merged instancee. Only if merge_classes is True. Default is "mix".
+
+    Raises
+    ------
+        AssertionError
+            One of the labels is not in ["Seg", "bxox_2d"].
+        AssertionError
+            One of the cameras is not in ["LV", "FV", "MVL", "MVR"].
+        AssertionError
+            One of the passed clases is not available/supported.
+
+    """
+
+    CAMERAS = [
+        "RV",   # Right View
+        "FV",   # Front View
+        "MVL",  # Mirror Left View
+        "MVR",  # Mirror Right View
+        ]
+    LABELS = [
+        "seg",
+        "box_2d"
+        ]
+    SEG_CLASSES = [
+        "void",
+        "road",
+        "lanemarks",
+        "curb",
+        "person",
+        "rider",
+        "vehicles",
+        "bicycle",
+        "motorcycle",
+        "traffic_sign"
+        ]
+
+    def __init__(
+        self,
+        labels=[],
+        cameras=[],
+        fragment=0.9,
+        name="WoodScape",
+        seg_classes=[],
+        merge_classes=False,
+        rename_merged="mix",
+        **kwargs):
+        super().__init__(name=name, **kwargs)
+
+        cameras = self.CAMERAS if cameras == list() else cameras
+        labels = self.LABELS if labels == list() else labels
+
+        if isinstance(fragment , int):
+            pass
+        elif isinstance(fragment, float):
+            assert fragment <= 1 and fragment >= -1, "fragment of type float can not be higher than 1 of less than -1."
+        else:
+            raise AttributeError("Invalid type of fragment type.")
+
+        if seg_classes == list():
+            seg_classes = self.SEG_CLASSES
+
+        assert all([c in self.SEG_CLASSES for c in seg_classes]), f"some segmentation classes are invalid, supported classes are :\n {self.SEG_CLASSES}"
+        assert all([v in self.CAMERAS for v in cameras]), f"Some cameras are invalid, should be in {self.CAMERAS}"
+        assert all([l in self.LABELS for l in labels]), f"Some labels are invalid, should be ib {self.LABELS}"
+        assert isinstance(merge_classes, (int, bool)), "Invalid merge_classes argument"
+        assert isinstance(rename_merged, str), "Invalid rename_merged argument"
+
+        self.items = sorted(glob.glob(os.path.join(self.dataset_dir, "rgb_images", "*")))
+        self.items = self._filter_cameras(self.items, cameras) 
+        self.items = self._filter_non_png(self.items)
+        
+        self.labels = labels
+        self.cameras = cameras
+        self.seg_classes = seg_classes
+        self.merge_classes = merge_classes
+        self.num_seg_classes = len(seg_classes)
+        self.seg_classes_renamed = seg_classes if not merge_classes else [rename_merged]
+
+        # Encode fraction
+        self.fragment = min(abs(fragment), len(self)) if isinstance(fragment, int) else int(abs(fragment) * len(self))
+
+        # Restricting the number of samples
+        if fragment > 0:
+            self.items = self.items[:self.fragment]
+        else:
+            self.items = self.items[len(self) - self.fragment:]
+
+    def getitem(self, idx):
+        ipath = self.items[idx]
+        frame = Frame(ipath, names=tuple("CHW"))
+
+        if "seg" in self.labels:
+            segmentation = self._path2segLabel(ipath)
+            frame.append_segmentation(segmentation)
+        
+        if "box_2d" in self.labels:
+            _, H, W = frame.shape
+            bbox2d_path = self._path2boxLabel(ipath, frame_size=(H, W))
+            frame.append_boxes2d(bbox2d_path)
+        return frame
+
+    @staticmethod
+    def _path2segPath(path):
+        """Maps rgb image path to corresponding segmentation path
+        
+        Parameters
+        ----------
+            path: str 
+                Path to rgb image.
+
+        """
+        path, file = os.path.split(path)
+        path, _ = os.path.split(path)
+        path = os.path.join(path, "semantic_annotations", "gtLabels", file)
+        return path
+
+    @staticmethod
+    def _path2boxPath(path):
+        """Maps rgb image path to corresponding json 2dbbox file
+        
+        Parameters
+        ----------
+            path: str
+                path to rgb image   
+
+        """
+        path, file = os.path.split(path)
+        path, _ = os.path.split(path)
+        path = os.path.join(path, "box_2d_annotations", file.replace(".png", ".txt"))
+        return path
+
+    def _path2segLabel(self, path):
+        """Maps image path to segmentation mask
+        
+        Parametrs
+        ---------
+            path: str
+                path to rgb image
+
+        """
+        path = self._path2segPath(path)
+        mask = np.asarray(Image.open(path))
+        mask = self.mask_2d_idx_to_3d_onehot_mask(mask)
+        return mask
+
+    def mask_2d_idx_to_3d_onehot_mask(self, mask_2d):
+        """Converts 2d index encoding mask to 3d one hot encoding one
+        
+        Parameters
+        ----------
+            mask : np.ndarray
+                Mask of size (H, W) with int values
+        
+        """
+        sample_seg_classes = torch.unique(torch.Tensor(mask_2d.reshape(-1)))
+            
+        num_sample_seg_classes = len(self.seg_classes_renamed)
+        mask_3d = np.zeros((num_sample_seg_classes, ) + mask_2d.shape)
+
+        dec = 0
+        for i, name in enumerate(self.seg_classes):
+            if i in sample_seg_classes:
+                mask_3d[i - dec] += (mask_2d == self.SEG_CLASSES.index(name)).astype(int)
+                if self.merge_classes:
+                    dec += 1
+            else:
+                dec += 1
+
+        mask_3d = Mask(mask_3d, names=tuple("CHW"))
+        mlabels = Labels(torch.arange(num_sample_seg_classes).to(torch.float32), labels_names=self.seg_classes_renamed, names=("N"), encoding="id")
+        mask_3d.append_labels(mlabels)
+        return mask_3d
+
+    def _path2boxLabel(self, path, frame_size):
+        """Maps image patgh to bbox2d label
+        
+        Parameters
+        ----------
+            path: str
+                rgb image path
+
+        """
+        path = self._path2boxPath(path)
+
+        with open(path, "r") as f:
+            content = f.readlines()
+        content = [x.replace("\n", "") for x in content]
+        content = [x.split(",") for x in content]
+        bboxs2d = [[int(x) for x in c[2:]] for c in content]
+        return BoundingBoxes2D(bboxs2d, boxes_format="xyxy", absolute=True, frame_size=frame_size)
+
+    @staticmethod
+    def _filter_non_png(items):
+        """Filters non png files from a list of paths to files
+        
+        Parameters
+        ----------
+            items : List[str]
+                list of paths to filter
+
+        """
+        return [p for p in items if p.endswith(".png")]
+
+    @staticmethod
+    def _filter_cameras(items, cameras):
+        """Filters paths by given cameras list
+        
+        Parameters
+        ----------
+
+            cameras : List[str]
+                List of cameras
+
+        """
+        return list(filter(lambda x : any([v in x for v in cameras]), items))
+
+
+if __name__ == "__main__":
+    ds =  WooodScapeDataset(
+        labels=[],
+        cameras=[],
+        fragment=1.,
+        )
+    idx = 222
+    frame = ds[idx]
+    frame.get_view().render()

alodataset/woodScape_split_dataset.py

@@ -0,0 +1,21 @@
+from alodataset import Split, SplitMixin, WooodScapeDataset
+
+
+class WoodScapeSplitDataset(WooodScapeDataset, SplitMixin):
+    SPLIT_FOLDERS = {Split.VAL: -0.1, Split.TRAIN: 0.9}
+
+    def __init__(
+            self,
+            split=Split.TRAIN,
+            **kwargs
+            ):
+        self.split = split
+        super().__init__(fragment=self.get_split_folder(), **kwargs)
+
+
+if __name__ == "__main__":
+    val = WoodScapeSplitDataset(split=Split.VAL)
+    train = WoodScapeSplitDataset(split=Split.TRAIN)
+
+    print("val :",  len(val))
+    print("train :",  len(train))

alonet/__init__.py

@@ -10,5 +10,3 @@
 
 from . import detr_panoptic
 from . import deformable_detr_panoptic
-
-