neurogolf-solver / medal-solvers /TASK076_PROGRESS.md
rogermt's picture
Add Task 076 progress notes (202/266 best, key insights documented)
daad60a verified
|
Raw
History Blame
1.56 kB

Task 076 Progress

Best: 202/266 (v11 - 8-connected BFS for decoration detection)

What works:

  • 4-connected component detection for bars (scipy.ndimage.label)
  • 8-connected BFS from bars to find decorations (1, 2, 3 cells)
  • 8 transforms (4 rotations + 4 reflections) matching rotated template to incomplete bars
  • Using the 2-marker as the reference point for relative positioning
  • Applying rotated decorations at matching positions

What fails (64 examples):

  • ~20: No matching transform found (bar shape relative to 2 differs due to 2 being at different position along bar)
  • ~33: Template detection fails (BFS over-connects nearby objects)
  • ~11: Unknown

Key insight discovered:

The 2-marker can be at DIFFERENT positions along the bar in different objects of the same example. E.g., template has 2 at position 1 of 5 along bar, but incomplete has 2 at position 3 of 5. This means the relative coordinates of bar cells differ between template and incomplete even when the correct rotation is applied.

What to try next:

  1. Match bars by centroid-offset (align centroids, not 2-positions) for finding rotation
  2. Then compute the parallel offset (how much the 2 shifted along the bar after rotation)
  3. Apply that parallel offset to decorations (shift decorations along the bar axis)
  4. This requires knowing the bar axis direction after rotation to shift correctly

Files:

  • task076_solver_194.py (uploaded earlier, uses simple 8-transform on 4-adj decos)
  • v11 (202/266) is the best - uses 8-BFS from bar for decoration detection