correct the padding related calculation errors in SSIM

CHANGELOG.md

@@ -39,7 +39,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
--
+- Correct the padding related calculation errors in SSIM ([#2721](https://github.com/Lightning-AI/torchmetrics/pull/2721))
 
 
 ## [1.4.1] - 2024-08-02

src/torchmetrics/functional/image/_deprecated.py

@@ -168,7 +168,7 @@ def _multiscale_structural_similarity_index_measure(
     >>> preds = rand([3, 3, 256, 256])
     >>> target = preds * 0.75
     >>> _multiscale_structural_similarity_index_measure(preds, target, data_range=1.0)
-    tensor(0.9627)
+    tensor(0.9628)
 
     """
     _deprecated_root_import_func("multiscale_structural_similarity_index_measure", "image")

src/torchmetrics/functional/image/ssim.py

@@ -124,11 +124,15 @@ def _ssim_update(
     dtype = preds.dtype
     gauss_kernel_size = [int(3.5 * s + 0.5) * 2 + 1 for s in sigma]
 
-    pad_h = (gauss_kernel_size[0] - 1) // 2
-    pad_w = (gauss_kernel_size[1] - 1) // 2
+    if gaussian_kernel:
+        pad_h = (gauss_kernel_size[0] - 1) // 2
+        pad_w = (gauss_kernel_size[1] - 1) // 2
+    else:
+        pad_h = (kernel_size[0] - 1) // 2
+        pad_w = (kernel_size[1] - 1) // 2
 
     if is_3d:
-        pad_d = (gauss_kernel_size[2] - 1) // 2
+        pad_d = (kernel_size[2] - 1) // 2
         preds = _reflection_pad_3d(preds, pad_d, pad_w, pad_h)
         target = _reflection_pad_3d(target, pad_d, pad_w, pad_h)
         if gaussian_kernel:
@@ -164,25 +168,21 @@ def _ssim_update(
 
     ssim_idx_full_image = ((2 * mu_pred_target + c1) * upper) / ((mu_pred_sq + mu_target_sq + c1) * lower)
 
-    if is_3d:
-        ssim_idx = ssim_idx_full_image[..., pad_h:-pad_h, pad_w:-pad_w, pad_d:-pad_d]
-    else:
-        ssim_idx = ssim_idx_full_image[..., pad_h:-pad_h, pad_w:-pad_w]
-
     if return_contrast_sensitivity:
         contrast_sensitivity = upper / lower
         if is_3d:
             contrast_sensitivity = contrast_sensitivity[..., pad_h:-pad_h, pad_w:-pad_w, pad_d:-pad_d]
         else:
             contrast_sensitivity = contrast_sensitivity[..., pad_h:-pad_h, pad_w:-pad_w]
-        return ssim_idx.reshape(ssim_idx.shape[0], -1).mean(-1), contrast_sensitivity.reshape(
+
+        return ssim_idx_full_image.reshape(ssim_idx_full_image.shape[0], -1).mean(-1), contrast_sensitivity.reshape(
             contrast_sensitivity.shape[0], -1
         ).mean(-1)
 
     if return_full_image:
-        return ssim_idx.reshape(ssim_idx.shape[0], -1).mean(-1), ssim_idx_full_image
+        return ssim_idx_full_image.reshape(ssim_idx_full_image.shape[0], -1).mean(-1), ssim_idx_full_image
 
-    return ssim_idx.reshape(ssim_idx.shape[0], -1).mean(-1)
+    return ssim_idx_full_image.reshape(ssim_idx_full_image.shape[0], -1).mean(-1)
 
 
 def _ssim_compute(
@@ -507,7 +507,7 @@ def multiscale_structural_similarity_index_measure(
         >>> preds = rand([3, 3, 256, 256])
         >>> target = preds * 0.75
         >>> multiscale_structural_similarity_index_measure(preds, target, data_range=1.0)
-        tensor(0.9627)
+        tensor(0.9628)
 
     References:
         [1] Multi-Scale Structural Similarity For Image Quality Assessment by Zhou Wang, Eero P. Simoncelli and Alan C.

src/torchmetrics/image/_deprecated.py

@@ -44,7 +44,7 @@ class _MultiScaleStructuralSimilarityIndexMeasure(MultiScaleStructuralSimilarity
     >>> target = preds * 0.75
     >>> ms_ssim = _MultiScaleStructuralSimilarityIndexMeasure(data_range=1.0)
     >>> ms_ssim(preds, target)
-    tensor(0.9627)
+    tensor(0.9628)
 
     """
 

src/torchmetrics/image/ssim.py

@@ -274,7 +274,7 @@ class MultiScaleStructuralSimilarityIndexMeasure(Metric):
         >>> target = preds * 0.75
         >>> ms_ssim = MultiScaleStructuralSimilarityIndexMeasure(data_range=1.0)
         >>> ms_ssim(preds, target)
-        tensor(0.9627)
+        tensor(0.9628)
 
     """
 

tests/unittests/image/test_ssim.py

@@ -278,7 +278,16 @@ def test_ssim_invalid_inputs(pred, target, kernel, sigma, match):
         structural_similarity_index_measure(pred, target, kernel_size=kernel, sigma=sigma)
 
 
-def test_ssim_unequal_kernel_size():
+@pytest.mark.parametrize(
+    ("sigma", "kernel_size", "result"),
+    [
+        ((0.25, 0.5), None, torch.tensor(0.20977394)),
+        ((0.5, 0.25), None, torch.tensor(0.13884821)),
+        (None, (3, 5), torch.tensor(0.05032664)),
+        (None, (5, 3), torch.tensor(0.03472072)),
+    ],
+)
+def test_ssim_unequal_kernel_size(sigma, kernel_size, result):
     """Test the case where kernel_size[0] != kernel_size[1]."""
     preds = torch.tensor([
         [
@@ -306,24 +315,16 @@ def test_ssim_unequal_kernel_size():
             ]
         ]
     ])
-    # kernel order matters
-    assert torch.isclose(
-        structural_similarity_index_measure(preds, target, gaussian_kernel=True, sigma=(0.25, 0.5)),
-        torch.tensor(0.08869550),
-    )
-    assert not torch.isclose(
-        structural_similarity_index_measure(preds, target, gaussian_kernel=True, sigma=(0.5, 0.25)),
-        torch.tensor(0.08869550),
-    )
-
-    assert torch.isclose(
-        structural_similarity_index_measure(preds, target, gaussian_kernel=False, kernel_size=(3, 5)),
-        torch.tensor(0.05131844),
-    )
-    assert not torch.isclose(
-        structural_similarity_index_measure(preds, target, gaussian_kernel=False, kernel_size=(5, 3)),
-        torch.tensor(0.05131844),
-    )
+    if sigma is not None:
+        assert torch.isclose(
+            structural_similarity_index_measure(preds, target, gaussian_kernel=True, sigma=sigma),
+            result,
+        )
+    else:
+        assert torch.isclose(
+            structural_similarity_index_measure(preds, target, gaussian_kernel=False, kernel_size=kernel_size),
+            result,
+        )
 
 
 @pytest.mark.parametrize(
@@ -341,3 +342,19 @@ def test_full_image_output(preds, target):
     assert len(out) == 2
     assert out[0].numel() == 1
     assert out[1].shape == preds[0].shape
+
+
+def test_ssim_for_correct_padding():
+    """Check that padding is correctly added and removed for SSIM.
+
+    See issue: https://github.com/Lightning-AI/torchmetrics/issues/2718
+
+    """
+    preds = torch.rand([3, 3, 256, 256])
+    # let the edge of the image be 0
+    target = preds.clone()
+    target[:, :, 0, :] = 0
+    target[:, :, -1, :] = 0
+    target[:, :, :, 0] = 0
+    target[:, :, :, -1] = 0
+    assert structural_similarity_index_measure(preds, target) < 1.0