Bizarre results with inpaint_nans

Dear all,

I have

A={

 [4.1583]    [4.2132]    [1.4600e+004]    [100]     
    [4.0339]    [4.1117]    [1.4948e+004]    [100]     
    [3.9389]    [4.0431]    [1.9352e+004]    [100]     
    [3.9676]    [4.0655]    [1.5374e+004]    [100]     
    [3.9313]    [4.0153]    [1.5853e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.8720]    [3.9659]    [1.7952e+004]    [100]     
    [3.8829]    [4.0014]    [1.7192e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.8681]    [3.9610]    [1.7790e+004]    [100]     
    [4.3944]    [3.8585]    [1.7593e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.8086]    [3.8981]    [1.4981e+004]    [100]     
    [3.7495]    [3.8701]    [1.5271e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.6838]    [3.7952]    [1.5973e+004]    [100]     
    [3.7344]    [3.8012]    [1.6384e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.7861]    [3.8795]    [1.7397e+004]    [100]     
    [3.7793]    [3.8614]    [1.8571e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.6177]    [3.7094]    [1.9622e+004]    [100]     
    [4.2028]    [3.5645]    [1.9093e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.6123]    [3.6857]    [1.5312e+004]    [100]     
    [3.5771]    [3.6983]    [1.5883e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.6439]    [3.7417]    [1.6312e+004]    [100]     
    [3.6003]    [3.7064]    [1.6997e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.7250]    [3.8390]    [1.7667e+004]    [100]     
    [3.7227]    [3.8167]    [1.8742e+004]    [100]     
    [   NaN]    [   NaN]    [        NaN]    [NaN]     
    [3.6667]    [3.7456]    [1.8926e+004]    [100]     
    [3.9350]    [3.4062]    [1.8028e+004]    [ 95]    }

IF I apply

 out=inpaint_nans(cell2mat(A) ,2);

 I obtain 
 4.15830000000000  4.21320000000000  14600  100
4.03390000000000  4.11170000000000  14948  100
3.93890000000000  4.04310000000000  19352  100
3.96760000000000  4.06550000000000  15374  100
3.93130000000000  4.01530000000000  15853  100
3.90165000000000  2263.50209333333  9042.12552333333  100
3.87200000000000  3.96590000000000  17952  100
3.88290000000000  4.00140000000000  17192  100
3.87550000000000  2341.95677333333  9355.98919333333  100
3.86810000000000  3.96100000000000  17790  100
4.39440000000000  3.85850000000000  17593  100
4.10150000000000  2181.42882666667  8713.85720666667  100
3.80860000000000  3.89810000000000  14981  100
3.74950000000000  3.87010000000000  15271  100
3.71665000000000  2092.63518666667  8359.15879666667  100
3.68380000000000  3.79520000000000  15973  100
3.73440000000000  3.80120000000000  16384  100
3.76025000000000  2261.78425333333  9035.69606333333  100
3.78610000000000  3.87950000000000  17397  100
3.77930000000000  3.86140000000000  18571  100
3.69850000000000  2555.87181333333  10212.2179533333  100
3.61770000000000  3.70940000000000  19622  100
4.20280000000000  3.56450000000000  19093  100
3.90755000000000  2303.30873333333  9202.07718333334  100
3.61230000000000  3.68570000000000  15312  100
3.57710000000000  3.69830000000000  15883  100
3.61050000000000  2155.94680000000  8612.73670000000  100
3.64390000000000  3.74170000000000  16312  100
3.60030000000000  3.70640000000000  16997  100
3.66265000000000  2320.58881333333  9271.14720333333  100
3.72500000000000  3.83900000000000  17667  100
3.72270000000000  3.81670000000000  18742  100
3.69470000000000  2520.86853333333  10072.2171333333  100
3.66670000000000  3.74560000000000  18926  100
3.93500000000000  3.40620000000000  18028  95

If I apply

 out3=[];
for c = 1:size(A,2)
     out=inpaint_nans(cell2mat(A(:,c)),2);

out3=[out3       out ];

 end

I obtain

 4.15830000000000  4.21320000000000  14600  100
4.03390000000000  4.11170000000000  14948  100
3.93890000000000  4.04310000000000  19352  100
3.96760000000000  4.06550000000000  15374  100
3.93130000000000  4.01530000000000  15853  100
3.90165000000000  3.99060000000000  16902.5000000000  100
3.87200000000000  3.96590000000000  17952  100
3.88290000000000  4.00140000000000  17192  100
3.87550000000000  3.98120000000000  17491  100
3.86810000000000  3.96100000000000  17790  100
4.39440000000000  3.85850000000000  17593  100
4.10150000000000  3.87830000000000  16287  100
3.80860000000000  3.89810000000000  14981  100
3.74950000000000  3.87010000000000  15271  100
3.71665000000000  3.83265000000000  15622  100
3.68380000000000  3.79520000000000  15973  100
3.73440000000000  3.80120000000000  16384  100
3.76025000000000  3.84035000000000  16890.5000000000  100
3.78610000000000  3.87950000000000  17397  100
3.77930000000000  3.86140000000000  18571  100
3.69850000000000  3.78540000000000  19096.5000000000  100
3.61770000000000  3.70940000000000  19622  100
4.20280000000000  3.56450000000000  19093  100
3.90755000000000  3.62510000000000  17202.5000000000  100
3.61230000000000  3.68570000000000  15312  100
3.57710000000000  3.69830000000000  15883  100
3.61050000000000  3.72000000000000  16097.5000000000  100
3.64390000000000  3.74170000000000  16312  100
3.60030000000000  3.70640000000000  16997  100
3.66265000000000  3.77270000000000  17332  100
3.72500000000000  3.83900000000000  17667  100
3.72270000000000  3.81670000000000  18742  100
3.69470000000000  3.78115000000000  18834  100
3.66670000000000  3.74560000000000  18926  100
3.93500000000000  3.40620000000000  18028  95

AS you can see between these two outputs the 2nd and 3rd row are different. But the code was the same. Any explanation?

thanks

[Oleg Komarov]

Why don't you store your "numbers" in a double array from the beginning?

Check the different in size:

whos A
B = cell2mat(A)
whos B

Answer 1

[Oleg Komarov]

inpaint_nans() uses a 4 neighborood algorithm because it's designed for images. Thus, looping by column hides the two adjacent columns.

If you don't want to loop with inpaint_nans(), then you can always use interp1():

% Convert to double (again, store them as double from the beginnig)

B   = cell2mat(A);

% Values to interpolate

xi  = (1:size(A,1))';

% Index the non NaN

idx = ~isnan(B(:,1));

out = interp1(xi(idx),B(idx,:),xi);

[Sabbas]

So, If I understand correctly you mean that i can use interp1() if I do not want to loop. can I use inpaint_nans()without looping?.

[Oleg Komarov]

No, because it considers the 4 neighborood, i.e. values to the North, South, East, West.

[Sabbas]

ok. So inpaint_nans() WITH looping

[Oleg Komarov]

Just to be sure, but you are aware that the result is exactly the same.

In this case I would personally go for the interp1(), interpolation which inpaint_nans() does as well.

Why don't you like interp1?

[Sabbas]

because I do not know how to apply it when I have scattered NaNs. IN the above question if the first element of a row is NaN then the rest elements of that row are also NaN. HOw coould I modify the code

 idx = ~isnan(B(:,1));
out = interp1(xi(idx),B(idx,:),xi);

when my matrix has scattered NaNs?

thanks

[Oleg Komarov]

Then you have to loop with interp1 as well.

Answer 2

[Sean de Wolski]

The second and third row don't look any different to me:

4.03390000000000  4.11170000000000  14948  100
3.93890000000000  4.04310000000000  19352  100

versus

4.03390000000000  4.11170000000000  14948  100
3.93890000000000  4.04310000000000  19352  100

It's been awhile since I've used inpaint_nans but it doesn't surprise me that it would look at a 2d array differently than column vectors either.

Hopefully the master himself will chime in :)

[Sabbas]

Sorry I made a mistake.

look at the 6th and 7th row

 3.90165000000000  2263.50209333333  9042.12552333333  100
3.87200000000000  3.96590000000000  17952  100

 3.90165000000000  3.99060000000000  16902.5000000000  100
3.87200000000000  3.96590000000000  17952  100