{ "id": 410747, "title": "Weird samples?", "url": "/competitions/asl-fingerspelling/discussion/410747", "author_name": "Vadim Irtlach", "post_date": "2023-05-16T12:47:34.926000", "votes": 11, "comment_count": 7, "views": 0, "content": "

What are these samples? It looks very weird that there are only a few frames, but the labels are very long to show for such a small number of frames.

\n

\"Screenshot-4\"

", "messages": [ { "id": 2261593, "postDate": "2023-05-16T12:47:34.927Z", "content": "

What are these samples? It looks very weird that there are only a few frames, but the labels are very long to show for such a small number of frames.

\n

\"Screenshot-4\"

", "rawMarkdown": "What are these samples? It looks very weird that there are only a few frames, but the labels are very long to show for such a small number of frames.\n\n\"Screenshot-4\"", "votes": 11 }, { "id": 2261813, "postDate": "2023-05-16T15:06:55.530Z", "content": "

Please fix your image, it says \"image not found\"

", "rawMarkdown": "Please fix your image, it says \"image not found\"", "votes": 2, "replies": [ { "id": 2261829, "postDate": "2023-05-16T15:14:33.620Z", "content": "

Sorry, Mark. Here is the new image

\n

\"Screenshot-4\"

", "rawMarkdown": "Sorry, Mark. Here is the new image\n\n\"Screenshot-4\"", "votes": 1, "replies": [ { "id": 2262118, "postDate": "2023-05-16T18:34:30.787Z", "content": "

Thanks and sharp observation.

\n

There seem to be plenty of single frame samples labelled with long phrases as shown below.
\nThese are only recordings with a single frame from 30 parquet files, meaning there can be many more samples containing a few frames labelled, incorrectly, with long phrases.

\n

\"\"

\n

It might be an idea to exclude the samples from the first peak shown below, as it is it is unlikely phrases can be fingerspelled in a few frames.

\n

The general take away from your observation I believe is the dataset is quite noisy.

\n

I hope, and expect, the test set will contain a manually picked subset with only correctly labelled samples, but this should be confirmed by the competition host.

\n

\"\"

", "rawMarkdown": "Thanks and sharp observation.\n\nThere seem to be plenty of single frame samples labelled with long phrases as shown below.\nThese are only recordings with a single frame from 30 parquet files, meaning there can be many more samples containing a few frames labelled, incorrectly, with long phrases.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4433335%2F7e57ce00464d1b244087daec8b7c7768%2Ftemp.png?generation=1684261467580951&alt=media)\n\nIt might be an idea to exclude the samples from the first peak shown below, as it is it is unlikely phrases can be fingerspelled in a few frames.\n\nThe general take away from your observation I believe is the dataset is quite noisy.\n\nI hope, and expect, the test set will contain a manually picked subset with only correctly labelled samples, but this should be confirmed by the competition host.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4433335%2F724f6ab398a3f59dfc33e52b22e8fa99%2Fabc.png?generation=1684262126411883&alt=media)", "votes": 5 }, { "id": 2262141, "postDate": "2023-05-16T18:50:57.470Z", "content": "
There seem to be plenty of single frame samples labelled with long phrases as shown below.\n
\n

Yep, I expected this. Thank you for your \"deep\" and almost full (only 30 parquet files) analysis, I just did an analysis on the subset of data (~1000 samples from the first parquet files). I appreciate such community intersections.

\n
The general take away from your observation I believe is the dataset is quite noisy.\n
\n

Yes, of course. We need some time to check it out, but the first hypothesis: \"maybe something was missed during preparation? Bug in software?\"

\n
I hope, and expect, the test set will contain a manually picked subset with only correctly labeled samples, but this should be confirmed by the competition host.\n
\n

I hope so.

", "rawMarkdown": "```\nThere seem to be plenty of single frame samples labelled with long phrases as shown below.\n```\n\nYep, I expected this. Thank you for your \"deep\" and almost full (only 30 parquet files) analysis, I just did an analysis on the subset of data (~1000 samples from the first parquet files). I appreciate such community intersections. \n\n\n```\nThe general take away from your observation I believe is the dataset is quite noisy.\n```\n\nYes, of course. We need some time to check it out, but the first hypothesis: \"maybe something was missed during preparation? Bug in software?\"\n\n```\nI hope, and expect, the test set will contain a manually picked subset with only correctly labeled samples, but this should be confirmed by the competition host.\n```\n\nI hope so. ", "votes": 1 }, { "id": 2295935, "postDate": "2023-06-11T12:06:09.260Z", "content": "

@sohier
\nThere are more than 1000 frames of length = 6. Can we assume that such a strange data set is included in the test data set?
\n\"\"

", "rawMarkdown": "@sohier \nThere are more than 1000 frames of length = 6. Can we assume that such a strange data set is included in the test data set?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1146523%2Fce2ee3765184be1bbf33c5fbf49cf95b%2F2023-06-11%20210355.png?generation=1686485056328741&alt=media)", "votes": 2 }, { "id": 2296130, "postDate": "2023-06-11T15:06:06Z", "content": "

there are also samples with only 1 frame 👀

", "rawMarkdown": "there are also samples with only 1 frame 👀", "votes": 1 }, { "id": 2369334, "postDate": "2023-08-01T17:07:51.867Z", "content": "

Hey! Did you find an answer if the test dataset would have a similar pattern?

", "rawMarkdown": "Hey! Did you find an answer if the test dataset would have a similar pattern?" } ] } ] } ], "comments": [ { "id": 2261813, "author_name": "Mark Wijkhuizen", "author_url": "", "post_date": "2023-05-16T15:06:55.530000", "content": "

Please fix your image, it says \"image not found\"

", "votes": 2, "replies": [ { "id": 2261829, "author_name": "Vadim Irtlach", "author_url": "", "post_date": "2023-05-16T15:14:33.620000", "content": "

Sorry, Mark. Here is the new image

\n

\"Screenshot-4\"

", "votes": 1, "replies": [ { "id": 2262118, "author_name": "Mark Wijkhuizen", "author_url": "", "post_date": "2023-05-16T18:34:30.787000", "content": "

Thanks and sharp observation.

\n

There seem to be plenty of single frame samples labelled with long phrases as shown below.
\nThese are only recordings with a single frame from 30 parquet files, meaning there can be many more samples containing a few frames labelled, incorrectly, with long phrases.

\n

\"\"

\n

It might be an idea to exclude the samples from the first peak shown below, as it is it is unlikely phrases can be fingerspelled in a few frames.

\n

The general take away from your observation I believe is the dataset is quite noisy.

\n

I hope, and expect, the test set will contain a manually picked subset with only correctly labelled samples, but this should be confirmed by the competition host.

\n

\"\"

", "votes": 5, "replies": [] }, { "id": 2262141, "author_name": "Vadim Irtlach", "author_url": "", "post_date": "2023-05-16T18:50:57.470000", "content": "
There seem to be plenty of single frame samples labelled with long phrases as shown below.\n
\n

Yep, I expected this. Thank you for your \"deep\" and almost full (only 30 parquet files) analysis, I just did an analysis on the subset of data (~1000 samples from the first parquet files). I appreciate such community intersections.

\n
The general take away from your observation I believe is the dataset is quite noisy.\n
\n

Yes, of course. We need some time to check it out, but the first hypothesis: \"maybe something was missed during preparation? Bug in software?\"

\n
I hope, and expect, the test set will contain a manually picked subset with only correctly labeled samples, but this should be confirmed by the competition host.\n
\n

I hope so.

", "votes": 1, "replies": [] }, { "id": 2295935, "author_name": "kurupical", "author_url": "", "post_date": "2023-06-11T12:06:09.260000", "content": "

@sohier
\nThere are more than 1000 frames of length = 6. Can we assume that such a strange data set is included in the test data set?
\n\"\"

", "votes": 2, "replies": [] }, { "id": 2296130, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2023-06-11T15:06:06", "content": "

there are also samples with only 1 frame 👀

", "votes": 1, "replies": [] }, { "id": 2369334, "author_name": "Quentin Parrenin", "author_url": "", "post_date": "2023-08-01T17:07:51.867000", "content": "

Hey! Did you find an answer if the test dataset would have a similar pattern?

", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2261593": "What are these samples? It looks very weird that there are only a few frames, but the labels are very long to show for such a small number of frames.\n\n\"Screenshot-4\"", "2261813": "Please fix your image, it says \"image not found\"" } }