Skip crowd images

alodataset/coco_base_dataset.py

@@ -16,7 +16,6 @@
 from typing import Dict, Union
 
 
-
 class CocoBaseDataset(BaseDataset):
     """
     Attributes
@@ -73,22 +72,29 @@ def __init__(
         classes: list = None,
         fix_classes_len: int = None,
         return_multiple_labels: list = None,
+        ignore_classes: list = None,
+        skip_crowd: bool = True,
         **kwargs,
     ):
         super(CocoBaseDataset, self).__init__(name=name, **kwargs)
         if self.sample:
             return
         else:
             assert img_folder is not None, "When sample = False, img_folder must be given."
-            assert ann_file is not None or "test" in img_folder, "When sample = False and the test split is not used, ann_file must be given."
-
+            assert (
+                ann_file is not None or "test" in img_folder
+            ), "When sample = False and the test split is not used, ann_file must be given."
 
         # Create properties
         self.img_folder = os.path.join(self.dataset_dir, img_folder)
 
         if "test" in img_folder:
-            #get a list of  indices that don't rely on the annotation file
-            self.items = [int(Path(os.path.join(self.img_folder, f)).stem) for f in os.listdir(self.img_folder) if os.path.isfile(os.path.join(self.img_folder, f))]
+            # get a list of  indices that don't rely on the annotation file
+            self.items = [
+                int(Path(os.path.join(self.img_folder, f)).stem)
+                for f in os.listdir(self.img_folder)
+                if os.path.isfile(os.path.join(self.img_folder, f))
+            ]
             return
 
         self.coco = COCO(os.path.join(self.dataset_dir, ann_file))
@@ -103,6 +109,12 @@ def __init__(
 
         self._ids_renamed = classes
         self.label_names = label_names
+
+        if classes is not None and ignore_classes is not None:
+            raise Exception("Can't considere both classes & ignore_classes at the same time.")
+        if ignore_classes is not None:
+            classes = [l for l in self.label_names if l not in set(ignore_classes + ["N/A"])]
+
         if classes is not None:
             notclass = [label for label in classes if label not in self.label_names]
             if len(notclass) > 0:  # Ignore all labels not in classes
@@ -122,6 +134,13 @@ def __init__(
                     ids.append(i)
             self.items = ids  # Remove images without bboxes with classes in classes list
 
+        if skip_crowd:
+            # Check each annotation and remove crowded ones.
+            for i in range(len(self.items) - 1, -1, -1):
+                anns = self.coco.loadAnns(self.coco.getAnnIds(self.items[i]))
+                if any([ann["iscrowd"] for ann in anns]):
+                    del self.items[i]
+
         # Fix lenght of label_names to a desired `fix_classes_len`
         if fix_classes_len is not None:
             self._fix_classes(fix_classes_len)
@@ -172,7 +191,12 @@ def _fix_classes(self, new_label_size):
 
     def _append_labels(self, element: Union[BoundingBoxes2D, Mask], target):
         def append_new_labels(element, ltensor, lnames, name):
-            label_2d = Labels(ltensor.to(torch.float32), labels_names=lnames, names=("N"), encoding="id")
+            label_2d = Labels(
+                ltensor.to(torch.float32),
+                labels_names=lnames,
+                names=("N"),
+                encoding="id",
+            )
             element.append_labels(label_2d, name=name)
 
         labels = target["labels"]
@@ -186,7 +210,10 @@ def append_new_labels(element, ltensor, lnames, name):
         # Append supercategory labels
         for ktype in self.label_types:
             append_new_labels(
-                element, torch.as_tensor(self.label_types[ktype])[labels], self.label_types_names[ktype], ktype
+                element,
+                torch.as_tensor(self.label_types[ktype])[labels],
+                self.label_types_names[ktype],
+                ktype,
             )
 
         # Append specific labels
@@ -211,7 +238,11 @@ def _target2aloscene(self, target, frame):
 
         # Create and append labels to boxes
         boxes = BoundingBoxes2D(
-            target["boxes"], boxes_format="xyxy", absolute=True, frame_size=frame.HW, names=("N", None)
+            target["boxes"],
+            boxes_format="xyxy",
+            absolute=True,
+            frame_size=frame.HW,
+            names=("N", None),
         )
         self._append_labels(boxes, target)
 
@@ -241,7 +272,7 @@ def getitem(self, idx):
 
         image_id = self.items[idx]
         if "test" in self.img_folder:
-            #get the filename from image_id without relying on annotation file
+            # get the filename from image_id without relying on annotation file
             return Frame(os.path.join(self.img_folder, f"{str(image_id).zfill(12)}.jpg"))
 
         frame = Frame(os.path.join(self.img_folder, self.coco.loadImgs(image_id)[0]["file_name"]))
@@ -293,7 +324,6 @@ def convert_coco_poly_to_mask(self, segmentations, height, width):
         return masks
 
     def __call__(self, image, target):
-
         w, h = image.shape[-1], image.shape[-2]
 
         image_id = target["image_id"]
@@ -356,11 +386,11 @@ def __call__(self, image, target):
 
 if __name__ == "__main__":
     coco_dataset = CocoBaseDataset(sample=False, img_folder="test2017")
-    #checking if regular getitem works
+    # checking if regular getitem works
     frame = coco_dataset[0]
     frame.get_view().render()
 
-    #check if dataloader works
+    # check if dataloader works
     for f, frames in enumerate(coco_dataset.train_loader(batch_size=2)):
         frames = Frame.batch_list(frames)
         frames.get_view().render()

alodataset/coco_detection_dataset.py

@@ -42,6 +42,9 @@ class CocoDetectionDataset(CocoBaseDataset, SplitMixin):
         Return Labels as a dictionary, with all posible categories found in annotations file, by default False
     ann_file : str
         Start from a fixe given annotation file where the path is relative to the `dataset_dir`
+    skip_crowd : bool, optional
+        Filter out images with `iscrowd` attribute, by default False
+        Images with crowd are often mislabeled with missing boxes for some persons.
     **kwargs : dict
         :mod:`BaseDataset <base_dataset>` optional parameters
 
@@ -80,6 +83,7 @@ def __init__(
         fix_classes_len: int = None,
         split=Split.TRAIN,
         ann_file=None,
+        skip_crowd: bool = False,
         **kwargs,
     ):
         SplitMixin.__init__(self, split)
@@ -144,6 +148,13 @@ def __init__(
                 if fix_classes_len is not None:
                     self._fix_classes(fix_classes_len)
 
+            if skip_crowd:
+                # Check each annotation and remove crowded ones.
+                for i in range(len(self.items) - 1, -1, -1):
+                    target = self.items[i][2]["segments_info"]
+                    if any([seg["iscrowd"] for seg in target]):
+                        del self.items[i]
+
         # Re-calcule encoding label types (+stuff)
         if self.label_types is not None:
             dict_cats = dict()

alodataset/coco_panoptic_dataset.py

@@ -50,6 +50,9 @@ class CocoPanopticDataset(BaseDataset, SplitMixin):
     fix_classes_len : int, optional
         Fix to a specific number the number of classes, filling the rest with "N/A" value.
         Use when the number of model outputs does not match with the number of classes in the dataset, by default 250
+    skip_crowd : bool, optional
+        Filter out images with `iscrowd` attribute, by default False
+        Images with crowd are often mislabeled for person boxes and instance segmentation.
     **kwargs : dict
         :mod:`BaseDataset <base_dataset>` optional parameters
 
@@ -61,7 +64,10 @@ class CocoPanopticDataset(BaseDataset, SplitMixin):
     """
 
     SPLIT_FOLDERS = {Split.VAL: "val2017", Split.TRAIN: "train2017"}
-    SPLIT_ANN_FOLDERS = {Split.VAL: "annotations/panoptic_val2017", Split.TRAIN: "annotations/panoptic_train2017"}
+    SPLIT_ANN_FOLDERS = {
+        Split.VAL: "annotations/panoptic_val2017",
+        Split.TRAIN: "annotations/panoptic_train2017",
+    }
     SPLIT_ANN_FILES = {
         Split.VAL: "annotations/panoptic_val2017.json",
         Split.TRAIN: "annotations/panoptic_train2017.json",
@@ -73,9 +79,10 @@ def __init__(
         split=Split.TRAIN,
         return_masks: bool = True,
         classes: list = None,
-        ignore_classes: list=None,
-        fix_classes_len: int = None,  # Match with pre-trained weights
-        **kwargs,
+        ignore_classes: list = None,
+        fix_classes_len: int = None,
+        skip_crowd: bool = False,
+        **kwargs,  # Match with pre-trained weights
     ):
         super(CocoPanopticDataset, self).__init__(name=name, split=split, **kwargs)
         if self.sample:
@@ -99,7 +106,7 @@ def __init__(
         if classes is not None:
             if self.label_names is None:
                 raise Exception(
-                    "'classes' attribute not support in datasets without 'categories' as attribute in annotation file"
+                    "'classes' attribute not supported in datasets without 'categories' as attribute in annotation file"
                 )
             notclass = [label for label in classes if label not in self.label_names]
             if len(notclass) > 0:  # Ignore all labels not in classes
@@ -111,7 +118,7 @@ def __init__(
             self._ids_renamed = np.array(self._ids_renamed)
             self.label_names = classes
 
-            # Check each annotation and keep only that have at least 1 element in classes list
+            # Check each annotation and keep only that have at least 1 element in classes
             items = []
             for i, (_, _, ann_info) in enumerate(self.items):
                 target = ann_info["segments_info"]
@@ -137,6 +144,13 @@ def __init__(
                     self.label_types[ltype] = [x for _, x in sorted(zip(idx_sort, self.label_types[ltype]))]
             self._ids_renamed = torch.from_numpy(self._ids_renamed)
 
+        if skip_crowd:
+            # Check each annotation and remove crowded ones.
+            for i in range(len(self.items) - 1, -1, -1):
+                target = self.items[i][2]["segments_info"]
+                if any([seg["iscrowd"] for seg in target]):
+                    del self.items[i]
+
         # Fix number of label names if desired
         if fix_classes_len is not None:
             if fix_classes_len > len(self.label_names):
@@ -269,7 +283,12 @@ def getitem(self, idx):
 
         # Make aloscene.frame
         frame = Frame(img_path)
-        labels_2d = Labels(labels.to(torch.float32), labels_names=self.label_names, names=("N"), encoding="id")
+        labels_2d = Labels(
+            labels.to(torch.float32),
+            labels_names=self.label_names,
+            names=("N"),
+            encoding="id",
+        )
         boxes_2d = BoundingBoxes2D(
             masks_to_boxes(masks),
             boxes_format="xyxy",
@@ -291,15 +310,15 @@ def getitem(self, idx):
 
 if __name__ == "__main__":
     # coco_seg = CocoPanopticDataset(sample=True)
-    coco_seg = CocoPanopticDataset()  # test
-    for f, frames in enumerate(coco_seg.train_loader(batch_size=2)):
-        frames = Frame.batch_list(frames)
-        labels_set = "category" if isinstance(frames.boxes2d[0].labels, dict) else None
-        views = [fr.boxes2d.get_view(fr, labels_set=labels_set) for fr in frames]
-        if frames.segmentation is not None:
-            views += [fr.segmentation.get_view(fr, labels_set=labels_set) for fr in frames]
-        frames.get_view(views).render()
-        # frames.get_view(labels_set=labels_set).render()
-
-        if f > 1:
-            break
+    coco_seg = CocoPanopticDataset(classes=["person"], skip_crowd=False)  # test
+    for f, frames in enumerate(
+        coco_seg.train_loader(batch_size=1, num_workers=0, sampler=torch.utils.data.SequentialSampler(coco_seg))
+    ):
+        if f > 10:
+            frames = Frame.batch_list(frames)
+            labels_set = "category" if isinstance(frames.boxes2d[0].labels, dict) else None
+            views = [fr.boxes2d.get_view(fr, labels_set=labels_set) for fr in frames]
+            if frames.segmentation is not None:
+                views += [fr.segmentation.get_view(fr, labels_set=labels_set) for fr in frames]
+            frames.get_view(views).render()
+            # frames.get_view(labels_set=labels_set).render()