Files to be opened with ImageJ (https://imagej.nih.gov/ij/) Images are generated with GPU code running from Java program using JCuda https://git.elphel.com/Elphel/tile_processor_gpu/tree/separate-compilation , https://git.elphel.com/Elphel/imagej-elphel/tree/lwir-distort, Next images are calculated from 4 of 5Mpix images, with nominal disparity 56 pixels applied to each of the 324x242 tiles (tiles are 16x16 pixels on the 8x8 pixel grid) 1562390331_711112-SHIFTED-D56.0.tiff (57.8MB)- set of 4 images calculated for nominal dispartity of 56 pixels, ERS (electronic rolling shutter) and optical aberration correction applied. Rotating mouse wheel in ImageJ or pressing "play" button cycles through all 4 images 1562390331_711112-CORR2D.tiff (179.6MB) - central 9x9 areas (of 15x15) of the per-tile 2D phase correlations. Meaningfull results only for the tiles that have disparity close to selected 56 pixels. 6 slices correspond to 2 horizontal camera pairs, (top, bottom), 2 vertical (left, right), and 2 diagonal (main diagonal, other diagonal). These 2D correlations are fed to the DNN to extract depths. 1562390331_711112-RGBA-D56.0:0:0:324:242:R.tiff (14.4MB)- generated texture from combination of the 4 images. It is just a test, because normally only small almost flat areas are generated at once, with each tile having correct disparity. Here all the tiles are rendered with the same 56pix disparity so only tiles with that disparity produce usable data. Image is RGBA but it is not rendered as such in ImageJ 1562390331_711112-RGBA-STACK-D56.0:0:0:324:242:R.tiff (76.6MB) - same as previous but each channel R,G,B,A represented separately (Alpha has different scale, In ImageJ: Image->Adjust-> "Brightness and Contrast"->Auto. The higher the level (1.0 max), the better match between images is detected. So all white is where the tiles have exactly 56.0pix disparity, or where images do not have high-frequency components (e.g. in the blue sky)