diff --git a/modules/objdetect/src/hog.cpp b/modules/objdetect/src/hog.cpp
index 1c7c8f73cb..7194b068ff 100644
--- a/modules/objdetect/src/hog.cpp
+++ b/modules/objdetect/src/hog.cpp
@@ -723,7 +723,7 @@ void HOGCache::init(const HOGDescriptor* _descriptor,
     #if CV_SIMD128
         idx = v_float32x4(0.0f, 1.0f, 2.0f, 3.0f);
 
-        for (; j <= blockSize.height - 4; j += 4)
+        for (; j <= blockSize.width - 4; j += 4)
         {
             v_float32x4 t = v_sub(idx, _bw);
             t = v_mul(t, t);
diff --git a/modules/objdetect/test/test_cascadeandhog.cpp b/modules/objdetect/test/test_cascadeandhog.cpp
index 08c78f8ae5..ddec0e7b25 100644
--- a/modules/objdetect/test/test_cascadeandhog.cpp
+++ b/modules/objdetect/test/test_cascadeandhog.cpp
@@ -1364,4 +1364,22 @@ TEST(Objdetect_CascadeDetector, small_img)
     }
 }
 
+// See https://github.com/opencv/opencv/issues/23580
+// HOGDescriptor::compute() overflowed the gaussian weights buffer when the block
+// was taller than it was wide: the SIMD store loop for the width buffer (_dj) was
+// bounded by blockSize.height instead of blockSize.width. The block is chosen wide
+// enough that the buffer is heap allocated, so the overflow is a real out-of-bounds
+// write. "Done" is simply that compute() runs without crashing.
+TEST(Objdetect_HOGDescriptor, issue_23580_tall_block_no_overflow)
+{
+    Size winSize(272, 2048); // height > width, width > AutoBuffer stack size
+    HOGDescriptor hog(winSize, /*blockSize*/ winSize, /*blockStride*/ Size(8, 8),
+                      /*cellSize*/ Size(8, 8), /*nbins*/ 9);
+
+    Mat src(winSize, CV_8UC1, Scalar::all(0));
+    std::vector<float> descriptors;
+    ASSERT_NO_THROW(hog.compute(src, descriptors));
+    EXPECT_FALSE(descriptors.empty());
+}
+
 }} // namespace