feat(compressors): add OWLv2 bbox crop to HashPipeline and refactor image utilities

- Add crop_batch method to HashPipeline for cropping images using OWLv2 detection boxes
- Integrate crop_batch into pipeline forward pass (extract_hash and extract_features)
- Move extract_masked_region from compressors/proposal/utils.py to utils/image.py
- Add crop_image_by_bbox utility function in utils/image.py
- Update type annotations to use dict[str, Any] instead of bare dict
- Update .justfile to add memory server command
- Update marimo dependency to >=0.22.0
- Update nvidia CUDA package markers for platform compatibility

mini-nav/utils/__init__.py

@@ -4,7 +4,7 @@ from .feature_extractor import (
     extract_single_image_feature,
     infer_vector_dim,
 )
-from .image import numpy_to_pil
+from .image import crop_image_by_bbox, extract_masked_region, numpy_to_pil
 
 __all__ = [
     "get_device",
@@ -13,4 +13,6 @@ __all__ = [
     "extract_single_image_feature",
     "extract_batch_features",
     "numpy_to_pil",
+    "extract_masked_region",
+    "crop_image_by_bbox",
 ]

mini-nav/utils/image.py

@@ -1,12 +1,64 @@
-"""Image conversion utilities."""
+"""Image utilities for conversion, masking, and cropping."""
 
 from __future__ import annotations
 
+from collections.abc import Sequence
+
 import numpy as np
+from numpy.typing import NDArray
 from PIL import Image
 
 
-def numpy_to_pil(rgb: np.ndarray) -> Image.Image:
+def extract_masked_region(
+    image: Image.Image,
+    mask: NDArray[np.bool_],
+) -> Image.Image:
+    """Extract masked region from image.
+
+    Args:
+        image: Original PIL Image.
+        mask: Binary mask as numpy array (True = keep).
+
+    Returns:
+        PIL Image with only the masked region visible.
+    """
+    image_np = np.array(image.convert("RGB"))
+    masked_np = image_np * mask[:, :, np.newaxis]
+    return Image.fromarray(masked_np.astype(np.uint8))
+
+
+def crop_image_by_bbox(
+    image: Image.Image,
+    bbox: Sequence[float],
+) -> Image.Image:
+    """Crop an image by bounding box [x1, y1, x2, y2].
+
+    Args:
+        image: Source PIL image.
+        bbox: OWLv2-style box coordinates [x1, y1, x2, y2].
+
+    Returns:
+        Cropped PIL image. Returns the original image when bbox is invalid.
+    """
+    if len(bbox) != 4:
+        return image
+
+    x1, y1, x2, y2 = tuple(float(v) for v in bbox)
+    if not np.isfinite([x1, y1, x2, y2]).all():
+        return image
+
+    left = max(0, int(np.floor(x1)))
+    top = max(0, int(np.floor(y1)))
+    right = min(image.width, int(np.ceil(x2)))
+    bottom = min(image.height, int(np.ceil(y2)))
+
+    if right <= left or bottom <= top:
+        return image
+
+    return image.crop((left, top, right, bottom))
+
+
+def numpy_to_pil(rgb: NDArray[np.uint8]) -> Image.Image:
     """Convert an RGB numpy array to a PIL Image.
 
     Handles arrays with 4 channels (RGBA) by dropping the alpha channel.