\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11704143%2Fb43be302473c9b8b0ab7c1ba84c46208%2F_20230723161334.png?generation=1690100033742842&alt=media\")
The correlation between the length of phrase label and the number of data frame should be positive, however we found in some cases only 2 or 3 frames of fingerspelling data means dozens of characters.
\nHere is the chart of frame length / phrase length in TRAINSET. There are so many samples that only use very few frames means so many characters.
\n
Here is an example which means \"request cherry\", and corresponding data as follow.![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11704143%2F39d831750a5eeb3428cf08590bc3fd28%2FGIF%202023-7-23%2016-24-28.gif?generation=1690100765463812&alt=media\")
There are a lot of samples with frame length / phrase length ≈ 0, not only in trainset, but also in valset. The reason for this problem may because of lost tracking of hand key points. So what should we do about these samples?
", "messages": [ { "id": 2355312, "postDate": "2023-07-23T08:43:57.247Z", "content": "The correlation between the length of phrase label and the number of data frame should be positive, however we found in some cases only 2 or 3 frames of fingerspelling data means dozens of characters.
\nHere is the chart of frame length / phrase length in TRAINSET. There are so many samples that only use very few frames means so many characters.
\n
Here is an example which means \"request cherry\", and corresponding data as follow.
There are a lot of samples with frame length / phrase length ≈ 0, not only in trainset, but also in valset. The reason for this problem may because of lost tracking of hand key points. So what should we do about these samples?
", "rawMarkdown": "The correlation between the length of phrase label and the number of data frame should be positive, however we found in some cases only 2 or 3 frames of fingerspelling data means dozens of characters.\n\nHere is the chart of frame length / phrase length in TRAINSET. There are so many samples that only use very few frames means so many characters.\n\n\nHere is an example which means \"request cherry\", and corresponding data as follow.\n\nThere are a lot of samples with frame length / phrase length ≈ 0, not only in trainset, but also in valset. The reason for this problem may because of lost tracking of hand key points. So what should we do about these samples?\n", "votes": 5 }, { "id": 2359834, "postDate": "2023-07-26T11:47:55.160Z", "content": "Maybe \"padding\" techniques can be applied here, which means extend short-length frames to regular-length frames, then do the prediction work.
", "rawMarkdown": "Maybe \"padding\" techniques can be applied here, which means extend short-length frames to regular-length frames, then do the prediction work.", "votes": 2 }, { "id": 2374008, "postDate": "2023-08-04T15:21:32.707Z", "content": "I feel the need to add one point: There are a lot of samples with frame length / phrase length ≈ 0 but it doesn't mean that all of the them lost tracking of hand key point.
\nFor skilled sign language speakers, the language can even have abbreviations, just as we speak as ordinary people.
When an experienced signer is receiving fingerspelling, they look at the overall shape of the the word rather than each individual letter to understand it. Sometimes when a signer is producing a word quickly, like E-L-E-P-H-A-N-T they will focus on producing the first few letters and the last letters clearly but may muddle or drop letters in the middle (E-L-E-muddle-N-T). The receiver can disambiguate the sign pretty easily given the context.\nCheck this link below. I hope this helps you.
\nhttps://www.kaggle.com/competitions/asl-fingerspelling/discussion/416470
Can we generate our own data to replace the dirty part of the original data? I think it's possible to improve the model's performance if this is approved.
\nFor example:
\nhttps://www.kaggle.com/code/hengck23/generate-your-own-pose-data-from-video
\nhttps://www.youtube.com/watch?v=2vaANtHSgvk
Maybe \"padding\" techniques can be applied here, which means extend short-length frames to regular-length frames, then do the prediction work.
", "votes": 2, "replies": [] }, { "id": 2374008, "author_name": "", "author_url": "", "post_date": "2023-08-04T15:21:32.707000", "content": "I feel the need to add one point: There are a lot of samples with frame length / phrase length ≈ 0 but it doesn't mean that all of the them lost tracking of hand key point.
\nFor skilled sign language speakers, the language can even have abbreviations, just as we speak as ordinary people.
When an experienced signer is receiving fingerspelling, they look at the overall shape of the the word rather than each individual letter to understand it. Sometimes when a signer is producing a word quickly, like E-L-E-P-H-A-N-T they will focus on producing the first few letters and the last letters clearly but may muddle or drop letters in the middle (E-L-E-muddle-N-T). The receiver can disambiguate the sign pretty easily given the context.\nCheck this link below. I hope this helps you.
\nhttps://www.kaggle.com/competitions/asl-fingerspelling/discussion/416470
Can we generate our own data to replace the dirty part of the original data? I think it's possible to improve the model's performance if this is approved.
\nFor example:
\nhttps://www.kaggle.com/code/hengck23/generate-your-own-pose-data-from-video
\nhttps://www.youtube.com/watch?v=2vaANtHSgvk