Sample Submission & Scoring

Hi @team,

  1. The sample submission format has patch data, coordinates, and rotation, but the starting notebooks shows only to export the coordinates and rotation. What’s the right way to submit samples?

  2. The test dataset doesn’t have coordinate info, and there might be an empty gap as seen in the picture. How do we make sure the full 2D data matches the ground truth size and gets the perfect score?

Thank you and have nice holiday ahead!

Hi @leocd91

The sample submission format includes a key corresponding to the image and an index for the patch, rather than the specific patch data itself. The patch data are represented as numpy arrays, what you would find in the test/train dataset. Correspondingly, the starting notebook demonstrates how to prepare a JSON file that includes coordinates and rotation, in relation to the key for the image and the index for the patch, totally as you could see in sample_submission.json.
So, this approach is the correct method to prepare submission.
Kind regards,
Onward Team.

1 Like

Hi @leocd91

The scoring algorithm takes into account the possibility that the size of the reconstructed image may not perfectly match the ground truth image. Therefore, any loss of score due to size discrepancies in such cases will be minimal and should not significantly impact the overall score.

Kind regards,
Onward Team.

2 Likes

Hi Xeek, same with the Parallel perspective challenge, does the evaluation give the same score if the seismic is flipped upside down?

Hi @leocd91 you are correct, the evaluation does indeed give the same score if the seismic is flipped upside down.

Happy puzzling

Onward Team

1 Like