added support for even sized subsets for MULTIWINDOW

src/pyvale/dic/_dicchecks.py

@@ -87,9 +87,17 @@ def _multiwindow_init(subset_size: int,
                              f"{multiwindow_search_area[i]} at index {i}")
 
     overlap = [x * (1.0-multiwindow_overlap) for x in multiwindow_subset_size]
+    configured_overlap = list(map(int, overlap))
     overlap.append(subset_step)
     overlap  = list(map(int,overlap))
 
+    # Avoid running the same final FFT level twice when size and overlap are identical
+    if (multiwindow_subset_size and
+            multiwindow_subset_size[-1] == subset_size and
+            multiwindow_search_area[-1] == subset_size and
+            configured_overlap[-1] == subset_step):
+        return configured_overlap, multiwindow_subset_size, multiwindow_search_area
+
     multiwindow_subset_size.append(subset_size)
     multiwindow_search_area.append(subset_size)
 
@@ -245,7 +253,9 @@ def _check_thresholds(threshold: float,
         raise ValueError("Optimizer precision must be a float strictly "
                          "between 0 and 1.")
 
-def _check_subsets(subset_size: int, subset_step: int) -> None:
+def _check_subsets(subset_size: int,
+                   subset_step: int,
+                   method: str | None = None) -> None:
     """
 
     Parameters
@@ -254,6 +264,9 @@ def _check_subsets(subset_size: int, subset_step: int) -> None:
         Threshold for the Levenberg optimization method.
     subset_step : int
         Threshold for the brute-force optimization method.
+    method : str, optional
+        Scan method. Even subset sizes are supported by the FFT-only
+        ``MULTIWINDOW`` method, but remain disallowed for the other methods.
 
     Raises
     ------
@@ -262,8 +275,8 @@ def _check_subsets(subset_size: int, subset_step: int) -> None:
     """
 
 
-    # Enforce scalar types for non-FFT methods
-    if subset_size % 2 == 0:
+    # The FFT-only multiwindow path handles even-sized subsets explicitly.
+    if subset_size % 2 == 0 and method != "MULTIWINDOW":
         raise ValueError("subset_size must be an odd number.")
 
     # check if subset_step is larger than the subset_size

src/pyvale/dic/dic2d.py

@@ -223,7 +223,7 @@ def calculate_2d(reference: np.ndarray | str | Path,
 
     dicchecks._check_thresholds(threshold, precision)
     common_util.check_output_directory(str(output_basepath), output_prefix, print_level)
-    dicchecks._check_subsets(subset_size, subset_step)
+    dicchecks._check_subsets(subset_size, subset_step, method_enum.value)
     updated_seeds = dicchecks._check_and_update_rg_seed(seed, roi_mask, method_enum.value, w, h, subset_size, subset_step)
     num_params = dicchecks._check_shape_function(shape_function_enum)
 

src/pyvale/dic/dic3d.py

@@ -78,7 +78,9 @@ def calculate_3d(reference: list[np.ndarray] | list[str] | list[Path],
         the existing flat format ``[x0, y0, x1, y1, ...]`` or tuple format
         ``[(x0, y0), (x1, y1), ...]``. If the method is not RG, this will be ignored.
     subset_size : int, optional
-        Size of the square subset window in pixels (default: 21).
+        Size of the square subset window in pixels (default: 21). Must be odd
+        except for the FFT-only ``"MULTIWINDOW"`` method, which also supports
+        even sizes.
     subset_step : int, optional
         Step size between subset centers in pixels (default: 10).
     correlation_criteria : str, optional
@@ -245,7 +247,7 @@ def calculate_3d(reference: list[np.ndarray] | list[str] | list[Path],
     # checks on the config
     dicchecks._check_thresholds(threshold, precision)
     common_util.check_output_directory(str(output_basepath), output_prefix, print_level)
-    dicchecks._check_subsets(subset_size, subset_step)
+    dicchecks._check_subsets(subset_size, subset_step, method_enum.value)
     updated_seeds = dicchecks._check_and_update_rg_seed(seed, roi_mask, method_enum.value, w0, h0, subset_size, subset_step)
     num_params = dicchecks._check_shape_function(shape_function_enum)
 

tests/dic/test_dic.py

@@ -466,8 +466,8 @@ def test_2d_multiwindow_fft_large(dic_temp_dir: Path) -> None:
         deformed=def_large,
         roi_mask=roi_rigid.mask,
         seed=seed,
-        subset_size=31,
-        subset_step=15,
+        subset_size=32,
+        subset_step=16,
         max_displacement=100,
         correlation_criteria="ZNSSD",
         shape_function="RIGID",