Adjust to favor the left input pixel for nearest-neighbor interpolation

This fixes a subtle issue where a resampled pixel falling exactly
between two input pixels would normally round half up, but would also be
rendered with rounding half up, resulting in the resampled pixel
having the wrong value.  The code now functionally rounds half down,
which cancels out the effect of rendering rounding half up in these
situations, and otherwise makes no difference.

Author: ayshih

lib/matplotlib/image.py

@@ -206,12 +206,24 @@ def _resample(
     out = np.zeros(out_shape + data.shape[2:], data.dtype)  # 2D->2D, 3D->3D.
     if resample is None:
         resample = image_obj.get_resample()
+
+    # When an output pixel falls exactly on the edge between two input pixels, the Agg
+    # resampler will use the right input pixel as the nearest neighbor.  We want the
+    # left input pixel to be chosen instead, so we flip the supplied transform.
+    if interpolation == 'nearest':
+        transform += Affine2D().translate(-out.shape[1], -out.shape[0]).scale(-1, -1)
+
     _image.resample(data, out, transform,
                     _interpd_[interpolation],
                     resample,
                     alpha,
                     image_obj.get_filternorm(),
                     image_obj.get_filterrad())
+
+    # Because we flipped the supplied transform, we then flip the output image back.
+    if interpolation == 'nearest':
+        out = np.flip(out, axis=(0, 1))
+
     return out