Z7, Nikon's next move

[Emmett]

Hi,

I have been working on sensor data (Bill Claff's Heat-map data) using multivariate statistics by taking all variables at the same time a while ago. Today I posted a short article regrading the Z7's sensor similarities to the current available camera sensors in the market. I hope it's useful for the enthusiasts :

https://multianalytics.blog/2018/10/29/adding-the-nikon-z7-to-the-equilibrium-nikon-next-move-on-sensor-properties-showed-to-be-as-expected/

[kers]

OK what data is on the x and y-axis?
unclear to me.

Pk

[mcbroomf]

What is the difference between T1/T2 and PC-1/PC-2 of the original article? Despite the overall similar "pattern" of camera clusters, when I overlay the graphs I don't see similar values, and also some cameras have moved relative to another, eg the Canon 1DxII vs the Sony A7s

[Dave Rosser]

OK what data is on the x and y-axis?
unclear to me.

Pk

+1

[Emmett]

OK what data is on the x and y-axis?
unclear to me.

Pk

X and Y are just coordinates. In other words in a PCA multivariate system each sensor will be given a score based on the sensor variables' values. They will place in a X,Y coordinate. For example X axis helps to maximize the PDR variability among sensors which in this case the more sensors are to the right the better their PDR and vice versa. QE is diagonal in this case, its lowest values starts from top left to the highest values on the bottom right and so on. You can even use this figure as a guide for sensor similarity and differences.
 If you read my previous article, it can help, but I am preparing simpler figures to clear things out.
https://multianalytics.blog/2018/01/31/multivariate-analysis-on-image-sensor-classification-and-variability-a-case-study-on-canon-nikon-and-sony/
 Check also the figure and explanation from Thom Hogan, I think he is extremely good at talking things simple and comprehensive:

http://www.dslrbodies.com/newsviews/an-intriguing-analysis.html
 

[Emmett]

What is the difference between T1/T2 and PC-1/PC-2 of the original article? Despite the overall similar "pattern" of camera clusters, when I overlay the graphs I don't see similar values, and also some cameras have moved relative to another, eg the Canon 1DxII vs the Sony A7s

Thanks for the question. I have removed DSNU and PRNU variables in this new PCA model and as you might have noticed the cumulative PCA score of the first two components has improved by 12.7% to 82.7%. In fact those two variables were making the model little unstable.

[mcbroomf]

Thanks for the question. I have removed DSNU and PRNU variables in this new PCA model and as you might have noticed the cumulative PCA score of the first two components has improved by 12.7% to 82.7%. In fact those two variables were making the model little unstable.

Thanks ...

[Emmett]

I added a few more description in a new post today, I hope it clarifies more:

https://multianalytics.blog/2018/11/02/how-a-global-camera-sensor-map-helps-you-to-decide-which-camera-to-choose-using-supervised-multivariate-classification-method/

[sbay]

Seems like the PCA is mostly picking up differences due to the size of the sensor which is not that interesting. I'd try normalizing that out of whatever input variables you have or running it within a class of sensors (e.g. only full frame camera models).

[Emmett]

Seems like the PCA is mostly picking up differences due to the size of the sensor which is not that interesting. I'd try normalizing that out of whatever input variables you have or running it within a class of sensors (e.g. only full frame camera models).

That was an OPLS not a PCA! There are six other variables there and sensor size has its own leverage on the model like other variables but not as much as you imagined. They are all significant variables, you don't leave a significant variable out unless you are looking for a purpose I don't understand or ou would want to purposely bias your model or ignore something. Check the variables, I think you can read them on the bi-plot.

[faberryman]

That was an OPLS not a PCA! There are six other variables there and sensor size has its own leverage on the model like other variables but not as much as you imagined. They are all significant variables, you don't leave a significant variable out unless you are looking for a purpose I don't understand or ou would want to purposely bias your model or ignore something. Check the variables, I think you can read them on the bi-plot.

Well,each sensor size occupies a similarly shaped horizontal band on the vertical axis, the larger the sensor, the higher up the axis. I suspect that if you eliminated sensor size, all the sensors would end up in one horizontal band. Not sure what the resulting band would show.

[mcbroomf]

I added a few more description in a new post today, I hope it clarifies more:

https://multianalytics.blog/2018/11/02/how-a-global-camera-sensor-map-helps-you-to-decide-which-camera-to-choose-using-supervised-multivariate-classification-method/

Thanks, good update (to my eyes).

I noticed that you have the Canon 1Ds II mislabeled(coloured) as a crop sensor (APSC).  It has a FF sensor.  The 1D series (not 1Ds series) were cropped, but to 1.3 not 1.5/1.6

[Emmett]

Well,each sensor size occupies a similarly shaped horizontal band on the vertical axis, the larger the sensor, the higher up the axis. I suspect that if you eliminated sensor size, all the sensors would end up in one horizontal band. Not sure what the resulting band would show.

They will not end up to a horizontal band unless the other variables has no leverage which is not the case here. most of them as you can see has leverage on t(2) axis which makes a cumulative contribution to that space and therefore will not end up to a flat line. I will post and image without sensor size factor later to address this.  Check figures 2a and 2b in the other post here and you will notice how those variables' loadings are making the pattern as you see:

https://multianalytics.blog/2018/01/31/multivariate-analysis-on-image-sensor-classification-and-variability-a-case-study-on-canon-nikon-and-sony/


 Furthermore like I already mentioned in that article, by side by side comparison of sensors' dynamic range and noise on DPreview, you will also find the similarities in the model (the closer the points in the figure the more similar they are on their characteristics). In that sense I believe DXO sensor scores are something I cannot suggest as reliable. You will appreciate how this global map can help quickly picking up sensor similarities on their various critical aspects without having to go through several pages of comparisons and benchmarks. Unfortunately I was unable to get more sensor data from other sources to make an even more refined model; my emails to them were left unanswered!  The data that I used are actually Bill Claff's; he was the only one who kindly provided his data.

[Emmett]

Thanks, good update (to my eyes).

I noticed that you have the Canon 1Ds II mislabeled(coloured) as a crop sensor (APSC).  It has a FF sensor.  The 1D series (not 1Ds series) were cropped, but to 1.3 not 1.5/1.6

Thanks for the heads up Mike I will fix the color.