I saw this effect in multiple parquet files. In the media-pipe documentation (https://google.github.io/mediapipe/solutions/hands.html#multi_hand_landmarks) its written that the x and y coordinates will be normalized between 0 and 1, by the input image width and height.
\nI loaded this parquet file: train_landmark_files/26734/1000035562.parquet, and observed the below output where max value of x and y is more than 1:
\nShapes of x, y, z are: (543,) (543,) (543,)\nmin and max of x: 0.1792217493057251 1.181456446647644\nmin and max of y: 0.22510389983654022 2.046358585357666\nmin and max of z: -2.329070806503296 1.4482454061508179\nI saw this effect in multiple parquet files. In the media-pipe documentation (https://google.github.io/mediapipe/solutions/hands.html#multi_hand_landmarks) its written that the x and y coordinates will be normalized between 0 and 1, by the input image width and height.
\nI loaded this parquet file: train_landmark_files/26734/1000035562.parquet, and observed the below output where max value of x and y is more than 1:
\nShapes of x, y, z are: (543,) (543,) (543,)\nmin and max of x: 0.1792217493057251 1.181456446647644\nmin and max of y: 0.22510389983654022 2.046358585357666\nmin and max of z: -2.329070806503296 1.4482454061508179\n< 0 and > 1 might mean extrapolated data off-camera. (Especially for 'pose' landmarks?) But that's just speculation.
", "rawMarkdown": "< 0 and > 1 might mean extrapolated data off-camera. (Especially for 'pose' landmarks?) But that's just speculation.", "votes": null }, { "id": "2168870", "postDate": "03/04/2023 15:43:05", "content": "Thanks, extrapolated off-camera means predicted by the media-pose model ? So, something like regression model but since its not bounded, in few cases it predicts out of training data range ?
\nAlso, I found this on their github: https://github.com/google/mediapipe/issues/2414#issuecomment-907587575
\nno problem，because the input of blazepose is the affined image that centered on mid hip，so back to original image, x or y may be greater than 1\nBut to be honest, I am a bit new to these terms and I was not able to understand it properly.
", "rawMarkdown": "Thanks, extrapolated off-camera means predicted by the media-pose model ? So, something like regression model but since its not bounded, in few cases it predicts out of training data range ? \n\nAlso, I found this on their github: https://github.com/google/mediapipe/issues/2414#issuecomment-907587575\n```\nno problem，because the input of blazepose is the affined image that centered on mid hip，so back to original image, x or y may be greater than 1\n```\n\nBut to be honest, I am a bit new to these terms and I was not able to understand it properly.", "votes": null }, { "id": "2168964", "postDate": "03/04/2023 16:59:56", "content": "It's just my guess, I don't really know
", "rawMarkdown": "It's just my guess, I don't really know", "votes": null }, { "id": "2169062", "postDate": "03/04/2023 18:52:16", "content": "Got it, makes sense. Thanks for your time.
", "rawMarkdown": "Got it, makes sense. Thanks for your time.", "votes": null }, { "id": "2171542", "postDate": "03/06/2023 21:36:40", "content": "We used the raw mediapipe landmarks so any points that are outside of [0, 1] are Mediapipe artifacts. I believe those are relatively rare, but unfortunately I don't have any more information about the root causes.
", "rawMarkdown": "We used the raw mediapipe landmarks so any points that are outside of [0, 1] are Mediapipe artifacts. I believe those are relatively rare, but unfortunately I don't have any more information about the root causes.", "votes": null }, { "id": "2171814", "postDate": "03/07/2023 05:36:51", "content": "Got it, thanks.
", "rawMarkdown": "Got it, thanks.", "votes": null } ], "comments": [ { "id": 2168836, "author_name": "roberthatch", "author_url": "", "post_date": "03/04/2023 15:13:19", "content": "< 0 and > 1 might mean extrapolated data off-camera. (Especially for 'pose' landmarks?) But that's just speculation.
", "votes": null, "replies": [ { "id": 2168870, "author_name": "yogendrayatnalkar", "author_url": "", "post_date": "03/04/2023 15:43:05", "content": "Thanks, extrapolated off-camera means predicted by the media-pose model ? So, something like regression model but since its not bounded, in few cases it predicts out of training data range ?
\nAlso, I found this on their github: https://github.com/google/mediapipe/issues/2414#issuecomment-907587575
\nno problem，because the input of blazepose is the affined image that centered on mid hip，so back to original image, x or y may be greater than 1\nBut to be honest, I am a bit new to these terms and I was not able to understand it properly.
", "votes": null, "replies": [ { "id": 2168964, "author_name": "roberthatch", "author_url": "", "post_date": "03/04/2023 16:59:56", "content": "It's just my guess, I don't really know
", "votes": null, "replies": [ { "id": 2169062, "author_name": "yogendrayatnalkar", "author_url": "", "post_date": "03/04/2023 18:52:16", "content": "Got it, makes sense. Thanks for your time.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2171542, "author_name": "sohier", "author_url": "", "post_date": "03/06/2023 21:36:40", "content": "We used the raw mediapipe landmarks so any points that are outside of [0, 1] are Mediapipe artifacts. I believe those are relatively rare, but unfortunately I don't have any more information about the root causes.
", "votes": null, "replies": [ { "id": 2171814, "author_name": "yogendrayatnalkar", "author_url": "", "post_date": "03/07/2023 05:36:51", "content": "Got it, thanks.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2168810": "I saw this effect in multiple parquet files. In the media-pipe documentation (https://google.github.io/mediapipe/solutions/hands.html#multi_hand_landmarks) its written that the x and y coordinates will be normalized between 0 and 1, by the input image width and height.\n\nI loaded this parquet file: **train_landmark_files/26734/1000035562.parquet**, and observed the below output where max value of x and y is more than 1:\t\n```\nShapes of x, y, z are: (543,) (543,) (543,)\nmin and max of x: 0.1792217493057251 1.181456446647644\nmin and max of y: 0.22510389983654022 2.046358585357666\nmin and max of z: -2.329070806503296 1.4482454061508179\n```", "2168836": "< 0 and > 1 might mean extrapolated data off-camera. (Especially for 'pose' landmarks?) But that's just speculation.", "2168870": "Thanks, extrapolated off-camera means predicted by the media-pose model ? So, something like regression model but since its not bounded, in few cases it predicts out of training data range ? \n\nAlso, I found this on their github: https://github.com/google/mediapipe/issues/2414#issuecomment-907587575\n```\nno problem，because the input of blazepose is the affined image that centered on mid hip，so back to original image, x or y may be greater than 1\n```\n\nBut to be honest, I am a bit new to these terms and I was not able to understand it properly.", "2168964": "It's just my guess, I don't really know", "2169062": "Got it, makes sense. Thanks for your time.", "2171542": "We used the raw mediapipe landmarks so any points that are outside of [0, 1] are Mediapipe artifacts. I believe those are relatively rare, but unfortunately I don't have any more information about the root causes.", "2171814": "Got it, thanks." }, "source": "meta" }