Unsigned Overflow in Python kernelP()

[Arya-Fgmain]

Hi there!

For starters, thank you & congratulations for your amazing work on this paper. It is incredibly insightful and helpful.

I found an issue in one of the scripts, utils.kernelP(), where the input image tensor has dtype np.uint8. The feature transformations in this function lead to unsigned overflow, especially the last 8 elements of each transformed data point (excluding the 1 at the end). The reason for this is mentioned in this page in the NumPy docs, which mentions that their numerical types are fixed-width. Below are the first pixels of 00-02.JPG before & after feature transformation, with their respective dtypes:

With Overflow:

(00.JPG) [ 66 25 131] uint8 , [ 66 25 131 114 198 203 4 113 9 86 1] int64
(01.JPG) [139 188 246] uint8 , [139 188 246 20 146 168 121 16 100 56 1] int64
(02.JPG) [215 252 254] uint8 , [215 252 254 164 82 8 145 16 4 184 1] int64

Without Overflow:

(00.JPG) [ 66 25 131] int64 , [ 66 25 131 1650 8646 3275 4356 625 17161 216150 1] int64
(01.JPG) [139 188 246] int64 , [ 139 188 246 26132 34194 46248 19321 35344 60516 6428472 1] int64
(02.JPG) [215 252 254] int64 , [ 215 252 254 54180 54610 64008 46225 63504 64516 13761720 1] int64

To fix the overflow issue, I added I = I.astype(np.int64) to the function.

I compared the output on the example_images directory with & without overflow. Find a summary of the respective MSE & MAE measures, as well as their averages, below. These metrics are to show how different the values are with and without overflow.

(MSE, MAE) for 00_AWB.png: (36.63265609741211, 4.617997646331787)
(MSE, MAE) for 00_T.png: (27.950292587280273, 4.0814528465271)
(MSE, MAE) for 00_S.png: (54.84620666503906, 5.604139804840088)
(MSE, MAE) for 01_AWB.png: (34.73400115966797, 4.047080039978027)
(MSE, MAE) for 01_T.png: (1.7762006521224976, 0.9251388907432556)
(MSE, MAE) for 01_S.png: (34.41278076171875, 4.021535396575928)
(MSE, MAE) for 02_AWB.png: (34.76441192626953, 4.017171382904053)
(MSE, MAE) for 02_T.png: (34.98468780517578, 3.8246243000030518)
(MSE, MAE) for 02_S.png: (32.551353454589844, 3.45462703704834)
(MSE, MAE) for 03_AWB.png: (6.443303108215332, 1.6022756099700928)
(MSE, MAE) for 03_T.png: (25.568115234375, 3.4743590354919434)
(MSE, MAE) for 03_S.png: (8.789534568786621, 1.9237173795700073)
(MSE, MAE) for 04_AWB.png: (19.582639694213867, 2.3370769023895264)
(MSE, MAE) for 04_T.png: (93.58589935302734, 5.58164119720459)
(MSE, MAE) for 04_S.png: (75.77192687988281, 4.169919013977051)
(MSE, MAE) for 05_AWB.png: (12.317625045776367, 2.122532606124878)
(MSE, MAE) for 05_T.png: (15.938911437988281, 2.534762382507324)
(MSE, MAE) for 05_S.png: (19.560606002807617, 2.326895236968994)
(MSE, MAE) for 06_AWB.png: (80.50083923339844, 6.207443714141846)
(MSE, MAE) for 06_T.png: (69.91217041015625, 6.178401470184326)
(MSE, MAE) for 06_S.png: (82.06550598144531, 6.34844446182251)
(MSE, MAE) for 07_AWB.png: (44.33774948120117, 4.364473342895508)
(MSE, MAE) for 07_T.png: (1.0995233058929443, 0.6863966584205627)
(MSE, MAE) for 07_S.png: (52.520164489746094, 4.696540832519531)
(MSE, MAE) for 08_AWB.png: (8.38996696472168, 1.8781654834747314)
(MSE, MAE) for 08_T.png: (11.96456241607666, 2.202268362045288)
(MSE, MAE) for 08_S.png: (27.575592041015625, 3.3244056701660156)
(MSE, MAE) for 09_AWB.png: (19.39771270751953, 2.680612564086914)
(MSE, MAE) for 09_T.png: (38.685218811035156, 3.597625732421875)
(MSE, MAE) for 09_S.png: (32.00599670410156, 3.630540132522583)
(MSE, MAE) for 10_AWB.png: (8.063882827758789, 1.7296009063720703)
(MSE, MAE) for 10_T.png: (13.180949211120605, 2.331162929534912)
(MSE, MAE) for 10_S.png: (10.85296630859375, 1.9405498504638672)
(MSE, MAE) for 11_AWB.png: (3.2594854831695557, 1.08352792263031)
(MSE, MAE) for 11_T.png: (98.14730834960938, 6.926990509033203)
(MSE, MAE) for 11_S.png: (9.633447647094727, 2.008676052093506)
(MSE, MAE) for 12_AWB.png: (4.093111515045166, 1.3582937717437744)
(MSE, MAE) for 12_T.png: (7.315742492675781, 1.7758421897888184)
(MSE, MAE) for 12_S.png: (7.722217559814453, 1.7602105140686035)

Mean MSE: 30.79321196
Mean MAE: 3.26607999

Please let me know what you think! I'd love to hear your POV.