Plotting a Pandas DataFrame with RGB values and coordinates

I have a pandas DataFrame with the columns ["x", "y", "r", "g", "b"] where x and y denote the coordinates of a pixel and r, g, b denote its RGB value. The rows contain entries for each coordinate of a grid of pixels and are unique. How can I display this DataFrame using matplotlibs's imshow()? This requires reshaping the data into a array of shape (M, N, 3).

My usual approach of using plt.imshow(df.pivot(columns="x", index="y", values="i"), interpolation="nearest") does only work for greyscale images. Placing ["r", "g", "b"] as the values argument yields a DataFrame with a MultiIndex as columns. However I fail to convert this into a correct image. Simply calling .reshape(M, N, 3) creates a wrong image.

I also had the idea of creating a new column with df["rgb"] = list(zip(df.r, df.g, df.b)) However I'm not sure on how to convert the resulting tuples into a new axis for the ndarray.

There exists an easy way to do this. First, you make sure the DataFrame is sorted by x- and y-values using df = df.sort_values(by=['x', 'y']).

Next, you select only the three columns for r, g and b from the DataFrame by calling df[['r', 'g', 'b']]. You convert the values into a numpy array by calling df[['r', 'g', 'b']].values, which will return an array of the shape (M*N, 3), assuming that M and N are the width and height of your image.
Now, reshape that array into the shape (M, N, 3) and you are done.

df = df.sort_values(by=['x', 'y'])
values = df[['r', 'g', 'b']].values
image = values.reshape(df['x'].max() + 1 , df['y'].max() + 1, 3)

I'm assuming here that your x and y values in the DataFrame start at 0, therefore I add 1 for the dimensions. If your x and y values start at 1, the reshaping can be done like this (df['x'].max(), df['y'].max(), 3).

Depending on what you consider the x and y dimensions of your image, you might have to transpose the array in the end.

Assuming this example:

   x  y    r    g    b
0  1  0    0  255  255
1  5  1  255  255    0
2  4  2  255    0  255
3  4  0    0  255  255
4  3  1  255  255    0
5  2  3  255    0    0

You could use:

N = df['x'].max()+1
M = df['y'].max()+1

tmp = (df.pivot(columns='x', index='y', values=['r', 'g', 'b'])
         .reindex(index=range(M),
                  columns=pd.MultiIndex.from_product([['r', 'g', 'b'],
                                                      range(N)
                                                     ]))
         .sort_index(level=1, axis=1, sort_remaining=False)
         .to_numpy().reshape((M, N, 3))
      )

plt.imshow(tmp, interpolation='nearest')

Output:

Alternatively, if you have all x/y:

np.random.seed(0)
M, N = 4, 5
df = pd.DataFrame({'x': np.repeat(np.arange(N), M),
                   'y': np.tile(np.arange(M), N),
                   'r': np.random.randint(0, 256, M*N),
                   'g': np.random.randint(0, 256, M*N),
                   'b': np.random.randint(0, 256, M*N),
                  })

tmp = (df[['x', 'y', 'r', 'g', 'b']]
 .melt(['x', 'y'], var_name='color')
 .assign(color=lambda d: pd.Categorical(d['color'],
                                        categories=['r', 'g', 'b'],
                                        ordered=True))
 .sort_values(by=['y', 'x', 'color'])
 ['value'].to_numpy()
 .reshape(M, N, 3)
)

plt.imshow(tmp, interpolation='nearest')

Output: