I watched this video - https://youtu.be/VKWxKE97Soc.
\nIt seems that for some signs you can use both hands in different order, like for the \"coffee\" sign. For signs with only one hand, you can use your left or right hand.
\nIt can be a problem if the training dataset consists only of right-handed signs and then we see left-handed signs.
\nAlso, how can we train the model because sometimes we can only see one hand and we can't pass landmarks for both hands.
\nMaybe we should use augmentation for this, like horizontal flipping, but for landmarks?
", "messages": [ { "id": "2196592", "postDate": "03/25/2023 14:21:30", "content": "I watched this video - https://youtu.be/VKWxKE97Soc.
\nIt seems that for some signs you can use both hands in different order, like for the \"coffee\" sign. For signs with only one hand, you can use your left or right hand.
\nIt can be a problem if the training dataset consists only of right-handed signs and then we see left-handed signs.
\nAlso, how can we train the model because sometimes we can only see one hand and we can't pass landmarks for both hands.
\nMaybe we should use augmentation for this, like horizontal flipping, but for landmarks?
", "rawMarkdown": "I watched this video - https://youtu.be/VKWxKE97Soc. \n\nIt seems that for some signs you can use both hands in different order, like for the \"coffee\" sign. For signs with only one hand, you can use your left or right hand. \n\nIt can be a problem if the training dataset consists only of right-handed signs and then we see left-handed signs. \n\nAlso, how can we train the model because sometimes we can only see one hand and we can't pass landmarks for both hands.\n\nMaybe we should use augmentation for this, like horizontal flipping, but for landmarks?", "votes": null }, { "id": "2196647", "postDate": "03/25/2023 15:00:35", "content": "I believe we are supposed to deal with one-hand signs -
\n\n", "rawMarkdown": "I believe we are supposed to deal with one-hand signs - \n\n> Signers pressed and held an on-screen button on the phone to record video while signing each concept, releasing the button after each sign.", "votes": null }, { "id": "2196767", "postDate": "03/25/2023 16:39:37", "content": "Signers pressed and held an on-screen button on the phone to record video while signing each concept, releasing the button after each sign.
\n
Tnx for your reply. My bad, didn't see this line.
", "rawMarkdown": "Tnx for your reply. My bad, didn't see this line.", "votes": null }, { "id": "2196822", "postDate": "03/25/2023 17:26:38", "content": "Nvm. That's the power of the kaggle community - helping each other 💪
", "rawMarkdown": "Nvm. That's the power of the kaggle community - helping each other 💪", "votes": null }, { "id": "2200908", "postDate": "03/28/2023 22:30:47", "content": "Each video should only have one hand signing, in some of the cases you may have both but that's because of some or the media pipe model extrapolating the landmarks (like it does for the pose landmarks). I would recommend you to try feature engineering the data by introducing new features that are not dependent on the coordinates location (e.g distance/angle between landmarks). This way you would not have to worry which hand is actually signing
", "rawMarkdown": "Each video should only have one hand signing, in some of the cases you may have both but that's because of some or the media pipe model extrapolating the landmarks (like it does for the pose landmarks). I would recommend you to try feature engineering the data by introducing new features that are not dependent on the coordinates location (e.g distance/angle between landmarks). This way you would not have to worry which hand is actually signing", "votes": null } ], "comments": [ { "id": 2196647, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "03/25/2023 15:00:35", "content": "I believe we are supposed to deal with one-hand signs -
\n\n", "votes": null, "replies": [ { "id": 2196767, "author_name": "pe4eniks", "author_url": "", "post_date": "03/25/2023 16:39:37", "content": "Signers pressed and held an on-screen button on the phone to record video while signing each concept, releasing the button after each sign.
\n
Tnx for your reply. My bad, didn't see this line.
", "votes": null, "replies": [ { "id": 2196822, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "03/25/2023 17:26:38", "content": "Nvm. That's the power of the kaggle community - helping each other 💪
", "votes": null, "replies": [] } ] } ] }, { "id": 2200908, "author_name": "josephzahar", "author_url": "", "post_date": "03/28/2023 22:30:47", "content": "Each video should only have one hand signing, in some of the cases you may have both but that's because of some or the media pipe model extrapolating the landmarks (like it does for the pose landmarks). I would recommend you to try feature engineering the data by introducing new features that are not dependent on the coordinates location (e.g distance/angle between landmarks). This way you would not have to worry which hand is actually signing
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2196592": "I watched this video - https://youtu.be/VKWxKE97Soc. \n\nIt seems that for some signs you can use both hands in different order, like for the \"coffee\" sign. For signs with only one hand, you can use your left or right hand. \n\nIt can be a problem if the training dataset consists only of right-handed signs and then we see left-handed signs. \n\nAlso, how can we train the model because sometimes we can only see one hand and we can't pass landmarks for both hands.\n\nMaybe we should use augmentation for this, like horizontal flipping, but for landmarks?", "2196647": "I believe we are supposed to deal with one-hand signs - \n\n> Signers pressed and held an on-screen button on the phone to record video while signing each concept, releasing the button after each sign.", "2196767": "Tnx for your reply. My bad, didn't see this line.", "2196822": "Nvm. That's the power of the kaggle community - helping each other 💪", "2200908": "Each video should only have one hand signing, in some of the cases you may have both but that's because of some or the media pipe model extrapolating the landmarks (like it does for the pose landmarks). I would recommend you to try feature engineering the data by introducing new features that are not dependent on the coordinates location (e.g distance/angle between landmarks). This way you would not have to worry which hand is actually signing" }, "source": "meta" }