dials.image_viewer: position of predictions seems to be wrong

[dagewa]

The display of predictions in the dials.image_viewer seems to be incorrect. For example, for i04-bag-training data, I index and find the spot on image 4 at around pixel (slow, fast) 1840,1835. I display the calculated centroid and the predicted centroid. Reading off values from the status bar I estimate the positions of the cross and the red spot as follows:

From image viewer
=================
            slow     fast
observed:   1840.785 1835.312
predicted:  1841.535 1835.812

Note that this implies there is a residual of about 0.9 pixels which seems rather high. Here are the images with values read off the status bar, highlighting where I had the cursor, as that was hidden while taking a screenshot:

Loading the reflection table in dials.python and selecting that reflection I can get at the real values for r['xyzobs.px.value'] and r['xyzcal.px']

>>> r['xyzobs.px.value']
(1835.2916666666667, 1840.75, 3.75)
>>> r['xyzcal.px']
(1835.289794544969, 1841.0089880829669, 3.7675501624336776)

So, the real residual on the image should only be 0.26 pixels, which better matches the RMSDs from indexing. Tidying up:

From reflection table
=====================
            slow     fast
observed:   1840.750 1835.292
predicted:  1841.009 1835.290

Comparing with the above table -- reminder:

From image viewer
=================
            slow     fast
observed:   1840.785 1835.312
predicted:  1841.535 1835.812

the observed position, i.e. the cross, seems to be about right whilst the prediction is way off.

[nksauter]

David, Just as a comment to this: when I originally wrote the viewer in Jan 2012 all the overlays were exact to numerical precision. I suspect the effect shown here arises from either rounding or failing to keep track of whether pixels have their center or corner at the registered position. No other insights. Nick Nicholas K. Sauter, Ph. D. Senior Scientist, Molecular Biophysics & Integrated Bioimaging Division Lawrence Berkeley National Laboratory 1 Cyclotron Rd., Bldg. 33R0345 Berkeley, CA 94720 (510) 486-5713

…

On Tue, Sep 12, 2017 at 9:16 AM, David Waterman ***@***.***> wrote: The display of predictions in the dials.image_viewer seems to be incorrect. For example, for i04-bag-training data, I index and find the spot on image 4 at around pixel (slow, fast) 1840,1835. I display the calculated centroid and the predicted centroid. Reading off values from the status bar I estimate the positions of the cross and the red spot as follows: From image viewer ================= slow fast observed: 1840.785 1835.312 predicted: 1841.535 1835.812 Note that this implies there is a residual of about 0.9 pixels which seems rather high. Here are the images with values read off the status bar, highlighting where I had the cursor, as that was hidden while taking a screenshot: [image: centroid] <https://user-images.githubusercontent.com/11502083/30336349-d10b1a6a-97dc-11e7-944d-99d99d57f972.png> [image: predicted] <https://user-images.githubusercontent.com/11502083/30336362-de1d15be-97dc-11e7-8c5b-40a571465d1b.png> Loading the reflection table in dials.python and selecting that reflection I can get at the real values for r['xyzobs.px.value'] and r['xyzcal.px'] >>> r['xyzobs.px.value'] (1835.2916666666667, 1840.75, 3.75) >>> r['xyzcal.px'] (1835.289794544969, 1841.0089880829669, 3.7675501624336776) So, the real residual on the image should only be 0.26 pixels, which better matches the RMSDs from indexing. Tidying up: From reflection table ===================== slow fast observed: 1840.750 1835.292 predicted: 1841.009 1835.290 Comparing with the above table -- reminder: From image viewer ================= slow fast observed: 1840.785 1835.312 predicted: 1841.535 1835.812 the observed position, i.e. the cross, seems to be about right whilst the prediction is way off. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#442>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AOGuVlwcHFLO5mxoc9k03ZKbc9MpAeTQks5shq5GgaJpZM4PU56I> .

[dagewa]

An easy way to investigate the issue is to set the predictions equal to the observations:

>>> from dials.array_family import flex
>>> rt=flex.reflection_table.from_pickle('indexed.pickle')
>>> rt['xyzcal.px']=rt['xyzobs.px.value']
>>> rt.as_pickle('pred_eq_obs.pickle')

Then all the spots show the same offset.

It does indeed look like a 0.5 pixel issue, in each direction.