Data - train, test and ground truth clarification

Hockey: Player localization
,
#1

Hi, I just want to clarify if my understanding of the data and expected output is correct.

  1. The final submission csv should contain the same 805 “videoId” and “frameId” combinations (rows) that are in “player_coords.csv”, just with the coordinates (“cords”) for the top-down rink view (“rink.jpg”), correct?
  2. Are the “cords” values in the 660 rows in “sample_submission.csv” the GROUND TRUTH for those “videoId” and “frameId” combinations that we can use to train our models?
  3. If #2. is correct, that means the 145 rows that are in “player_coords.csv” but not in “sample_submission.csv” represent the test set, correct?

Please confirm my understanding above or correct anything I’m misunderstanding. Thanks.

#2

Hi @justin_toribio

Thank you for reaching out;

  1. This is correct
  2. Sample submission is just a sample submission it has applied offset to the ground truth and has skipped some of the rows.
  3. There is currently no test data that includes the ground truth. I’ll be working on adding that to the resources this week.

Please let me know if you have any further questions

Cheers,

Jason

1 Like
#3

@justin_toribio I have added training data to https://www.aicrowd.com/challenges/hockey-player-localization/dataset_files

1 Like
#4

@jason_brumwell I’m quite confused about the coordinates. Is there any possible way to get the labelled coordinates of different positions like goal, blue line, etc which could be helpful for predicting the player coordinates?? or we just have to use train images and csv for training the model ignoring the top-down view?

#5

@Nischay My apologies this is something we do not currently have available. If you can provide a detailed spec of what you would require I can look at getting that done for you.

Are you looking for xy coordinates for goal posts, circles, line intersections with boards? Or are you looking for circles, rectangles etc?

#6

@jason_brumwell Basically, I wanted to know how these player coordinates were calculated. It would be helpful if you provide some details about the total range of xy coordinates for the board from goal to goal and coordinates of those circles, midline present in top view angle.

#7

@Nischay the ground truth was created using a custom manual homography tool. Some approaches that may be work researching are listed here How to get started?

1 Like