First Commit

requirements.txt

@@ -0,0 +1 @@
+datumaro

rotate_dataset.py

@@ -0,0 +1,89 @@
+import argparse
+from copy import deepcopy
+import datumaro as dm
+from datumaro.components.transformer import ItemTransform
+from datumaro.components.media import Image
+from datumaro.components.annotation import Bbox
+import numpy as np
+
+
+class RotateTransform(ItemTransform):
+    def __init__(self, extractor, angle):
+        super().__init__(extractor)
+        self.angle = angle
+        # k is number of 90-degree CCW rotations
+        self.k = (angle // 90) % 4
+
+    def transform_item(self, item):
+        if not item.media or not isinstance(item.media, Image):
+            return item
+
+        h, w = item.media.data.shape[:2]
+        # 1. Rotate Image
+        rotated_image = np.rot90(item.media.data, k=-self.k)  # CCW to CW adjustment
+
+        # 2. Rotate Bounding Boxes
+        new_annotations = []
+        for ann in item.annotations:
+            if isinstance(ann, Bbox):
+                x, y, bw, bh = ann.points
+                # Calculate new coordinates based on rotation angle
+                if self.angle == 90:
+                    new_bbox = [h - (y + bh), x, bh, bw]
+                elif self.angle == 180:
+                    new_bbox = [w - (x + bw), h - (y + bh), bw, bh]
+                elif self.angle == 270:
+                    new_bbox = [y, w - (x + bw), bh, bw]
+                else:
+                    new_bbox = [x, y, bw, bh]
+
+                new_annotations.append(ann.wrap(points=new_bbox))
+            else:
+                new_annotations.append(ann)
+
+        print(f"{item.id}_r{self.angle}")
+        return item.wrap(
+            id=f"{item.id}_r{self.angle}",
+            media=Image.from_numpy(rotated_image),
+            annotations=new_annotations,
+        )
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Rotate Datumaro datasets.")
+    parser.add_argument("--input", required=True, help="Path to input dataset")
+    parser.add_argument("--output", default="output_yolo", help="Output directory")
+    parser.add_argument(
+        "--angles",
+        nargs="+",
+        type=int,
+        default=[90, 180, 270],
+        help="Space separated angles (e.g. 90 180)",
+    )
+    args = parser.parse_args()
+
+    # 1. Load Original
+    dataset = dm.Dataset.import_from(args.input, "coco")
+    cloned_dataset = deepcopy(dataset)
+
+    # 2. Generate list of rotated datasets from args
+    rotated_sets = []
+    for angle in args.angles:
+        ds_rt = cloned_dataset.transform(RotateTransform, angle=angle)
+        rotated_sets.append(ds_rt)
+        cloned_dataset = deepcopy(dataset)
+
+    # rotated_sets = [dataset.transform(RotateTransform, angle=a) for a in args.angles]
+
+    # 3. Combine using splat operator on the array
+    combined_ds = dm.Dataset.from_extractors(dataset, *rotated_sets)
+    # combined_ds.transform("reindex", start=0)
+
+    # 4. Export to Ultralytics YOLO
+    # combined_ds.export(args.output, format="coco", save_media=True)
+    combined_ds.export(args.output, format="coco", save_media=True)
+    print(f"✅ Success! Augmented dataset saved to: {args.output}")
+
+
+if __name__ == "__main__":
+    main()