![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2F2389f248b65c749ef3f1f3e1df7112ee%2FScreenshot%20from%202023-06-28%2015-55-36.png?generation=1687982160870847&alt=media\")
Hi,
\nHaving tried to identify and verify duplicates, It did not take long to find this example:
\n
It looks like the same place but different day/time.
\nThis discovery has raised my concerns about potential data leakage and its impact on the model performance.
Further discussion Points:
\nHi,
\nHaving tried to identify and verify duplicates, It did not take long to find this example:
\n
It looks like the same place but different day/time.
\nThis discovery has raised my concerns about potential data leakage and its impact on the model performance.
Further discussion Points:
\nHey @sergiosaharovskiy,
\nTake a look at this:
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12870466%2F857328cdf8510888a36cb92eab022c2d%2FScreenshot_4.png?generation=1687984029852073&alt=media\")
Such duplicates are not hard to find when you look at the train and valid metadata provided by the organisers. I thus believe this in intentionnal since it would be an easy fix to make.
", "rawMarkdown": "Hey @sergiosaharovskiy, \nTake a look at this:\n\n\n\nSuch duplicates are not hard to find when you look at the train and valid metadata provided by the organisers. I thus believe this in intentionnal since it would be an easy fix to make. ", "votes": 4, "replies": [ { "id": 2321747, "postDate": "2023-06-28T20:47:11.243Z", "content": "It could be found by geo for sure, you are right! My concern is that mask overlaps as well…
", "rawMarkdown": "It could be found by geo for sure, you are right! My concern is that mask overlaps as well..." }, { "id": 2327577, "postDate": "2023-07-03T03:49:10.503Z", "content": "Yeah it is probably intentional since these photos are not exactly identical and taken 40 minutes apart (each record_id has 8 frames taken at 5 minute intervals), but I feel like the competition hosts overlooked how similar these two record_ids can be. Hopefully, the hidden test set has records taken a decent bit of time away from all frames in the train and validation folders.
", "rawMarkdown": "Yeah it is probably intentional since these photos are not exactly identical and taken 40 minutes apart (each record_id has 8 frames taken at 5 minute intervals), but I feel like the competition hosts overlooked how similar these two record_ids can be. Hopefully, the hidden test set has records taken a decent bit of time away from all frames in the train and validation folders. ", "votes": 2 } ] }, { "id": 2321771, "postDate": "2023-06-28T21:23:28.093Z", "content": "omg thanks for bringing this to notice
", "rawMarkdown": "omg thanks for bringing this to notice \n", "votes": 1 } ], "comments": [ { "id": 2321737, "author_name": "JEANMPIA", "author_url": "", "post_date": "2023-06-28T20:28:54.070000", "content": "Hey @sergiosaharovskiy,
\nTake a look at this:
Such duplicates are not hard to find when you look at the train and valid metadata provided by the organisers. I thus believe this in intentionnal since it would be an easy fix to make.
", "votes": 4, "replies": [ { "id": 2321747, "author_name": "SSS", "author_url": "", "post_date": "2023-06-28T20:47:11.243000", "content": "It could be found by geo for sure, you are right! My concern is that mask overlaps as well…
", "votes": 0, "replies": [] }, { "id": 2327577, "author_name": "Ari", "author_url": "", "post_date": "2023-07-03T03:49:10.503000", "content": "Yeah it is probably intentional since these photos are not exactly identical and taken 40 minutes apart (each record_id has 8 frames taken at 5 minute intervals), but I feel like the competition hosts overlooked how similar these two record_ids can be. Hopefully, the hidden test set has records taken a decent bit of time away from all frames in the train and validation folders.
", "votes": 2, "replies": [] } ] }, { "id": 2321771, "author_name": "Jayanth Sai Marreddi", "author_url": "", "post_date": "2023-06-28T21:23:28.093000", "content": "omg thanks for bringing this to notice
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "2321723": "Hi,\n\nHaving tried to identify and verify duplicates, It did not take long to find this example:\n\n\n\nIt looks like the same place but different day/time. \nThis discovery has raised my concerns about potential data leakage and its impact on the model performance.\n\nFurther discussion Points:\n\n1. Verifying more duplicates.\n2. How do we handle the duplicates?\n3. Fairness of the competition, can we expect leak in the private dataset?", "2321737": "Hey @sergiosaharovskiy, \nTake a look at this:\n\n\n\nSuch duplicates are not hard to find when you look at the train and valid metadata provided by the organisers. I thus believe this in intentionnal since it would be an easy fix to make. ", "2321771": "omg thanks for bringing this to notice \n" } }