Add puzzle keypoints.

README.md

@@ -26,5 +26,17 @@ The dataset is intended for training models that can automatically determine the
 ## Dataset Structure
 
 The dataset consists of the following fields:
-- `image` is an image of a Sudoku puzzle. The images are of varying sizes and the image formats include webp, jpg and png. Some images are screenshots and others are photos.
-- `cells` is an array representing the puzzle cells. It has shape (9, 9, 10) = (rows, cols, flags) and all entries are 0 or 1. Flag 0 encodes whether the cell is solved (1) or unsolved (0). Flags 1-9 encode whether the digits 1-9 respectively are present (1) or absent (0). If the cell is solved, then there is exactly one digit present (the solution) and all other digits are absent. If the cell is unsolved, then any set of digits can be present (including no digits). If the cell is unsolved, then any present digits are candidates, not solutions.
+
+- `image` is an image of a Sudoku puzzle.
+  
+  The images are of varying sizes and the image formats include webp, jpg and png. Some images are screenshots and others are photos.
+
+- `cells` is an array representing the puzzle cells.
+
+  It has shape (9, 9, 10) = (rows, cols, flags) and all entries are 0 or 1. Flag 0 encodes whether the cell is solved (1) or unsolved (0). Flags 1-9 encode whether the digits 1-9 respectively are present (1) or absent (0). If the cell is solved, then there is exactly one digit present (the solution) and all other digits are absent. If the cell is unsolved, then any set of digits can be present (including no digits). If the cell is unsolved, then any present digits are candidates, not solutions.
+
+- `keypoints` is an array of floats representing the puzzle keypoints (corners).
+
+  It has shape (8,) but represents four (x, y) coordinate pairs. The values are absolute image coordinates and start in the top-left of the image, so (0.0, 0.0) is the top-left and (width, height) is the bottom-right. The keypoints are all ordered in the same way: top-left, bottom-left, bottom-right, top-right.
+  
+  Note that for photos, the keypoints may be outside the puzzle corners. This is because in some photos, the puzzle is distorted, so performing a crop or perspective transform based on the real corner locations would exclude some of the puzzle. Having the keypoints slightly outside the puzzle corners in this case overcomes that problem.