Handle the overscan as integer to calculate the median

Even if the overscan image is floating, it's really ints with lots of
ties which must be handled carefully to get sub-integer precision.

You should never estimate overscan on "real" floating images, but
there's an option to prevent the conversion to int in case you do

python/lsst/ip/isr/isrFunctions.py

@@ -288,7 +288,8 @@ def illuminationCorrection(maskedImage, illumMaskedImage, illumScale):
     maskedImage.scaledDivides(1./illumScale, illumMaskedImage)
 
 
-def overscanCorrection(ampMaskedImage, overscanImage, fitType='MEDIAN', order=1, statControl=None):
+def overscanCorrection(ampMaskedImage, overscanImage, fitType='MEDIAN', order=1, statControl=None,
+                       overscanIsInt=True):
     """Apply overscan correction in-place
 
     The ``ampMaskedImage`` and ``overscanImage`` are modified, with the fit
@@ -319,6 +320,9 @@ def overscanCorrection(ampMaskedImage, overscanImage, fitType='MEDIAN', order=1,
         ``fitType`` indicates a polynomial or spline.
     statControl : `lsst.afw.math.StatisticsControl`
         Statistics control object.  In particular, we pay attention to numSigmaClip
+    overscanIsInt : `bool`
+        Treat the overscan region as consisting of integers, even if it's been
+        converted to float.  E.g. handle ties properly
 
     Returns
     -------
@@ -336,10 +340,27 @@ def overscanCorrection(ampMaskedImage, overscanImage, fitType='MEDIAN', order=1,
 
     numSigmaClip = statControl.getNumSigmaClip()
 
-    if fitType in ('MEAN', 'MEANCLIP', 'MEDIAN',):
+    if fitType in ('MEAN', 'MEANCLIP'):
         fitType = afwMath.stringToStatisticsProperty(fitType)
         offImage = afwMath.makeStatistics(overscanImage, fitType, statControl).getValue()
         overscanFit = offImage
+    elif fitType in ('MEDIAN',):
+        if overscanIsInt:
+            # we need an image with integer pixels to handle ties properly
+            if hasattr(overscanImage, "image"):
+                imageI = overscanImage.image.convertI()
+                overscanImageI = afwImage.MaskedImageI(imageI, overscanImage.mask, overscanImage.variance)
+            else:
+                overscanImageI = overscanImage.convertI()
+        else:
+            overscanImageI = overscanImage
+
+        fitType = afwMath.stringToStatisticsProperty(fitType)
+        offImage = afwMath.makeStatistics(overscanImageI, fitType, statControl).getValue()
+        overscanFit = offImage
+
+        if overscanIsInt:
+            del overscanImageI
     elif fitType in ('POLY', 'CHEB', 'LEG', 'NATURAL_SPLINE', 'CUBIC_SPLINE', 'AKIMA_SPLINE'):
         if hasattr(overscanImage, "getImage"):
             biasArray = overscanImage.getImage().getArray()