![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3700832%2F6531c5b0acbf90a46e7fdcedf944c4e6%2F2D_invis_obj.png?generation=1569940606872450&alt=media\")
\nIn sample annotation record there is a \"visibility_token\" field which I could use to filter out such object. At least, the guy from the nuscenes suggests to use it here. However, this field is empty in the Lyft dataset.I am trying to train own NN for 2D object detection on Lyft data. As the lyft dataset contains labels in 3D coordinates, we need to convert them into a camera image frame to get train data. \nIt is not a problem, however, there are some issues. The major issue that I cannot remove an invisible objects from labels.\nFor instance, as marked examples on image below\n\nIn sample annotation record there is a \"visibility_token\" field which I could use to filter out such object. At least, the guy from the nuscenes suggests to use it here. However, this field is empty in the Lyft dataset.
Is there a version of Lyft DS with this field?
", "messages": [ { "id": "638112", "postDate": "10/01/2019 14:44:27", "content": "I am trying to train own NN for 2D object detection on Lyft data. As the lyft dataset contains labels in 3D coordinates, we need to convert them into a camera image frame to get train data. \nIt is not a problem, however, there are some issues. The major issue that I cannot remove an invisible objects from labels.\nFor instance, as marked examples on image below\n\nIn sample annotation record there is a \"visibility_token\" field which I could use to filter out such object. At least, the guy from the nuscenes suggests to use it here. However, this field is empty in the Lyft dataset.
Is there a version of Lyft DS with this field?
", "rawMarkdown": "I am trying to train own NN for 2D object detection on Lyft data. As the lyft dataset contains labels in 3D coordinates, we need to convert them into a camera image frame to get train data. \nIt is not a problem, however, there are some issues. The major issue that I cannot remove an invisible objects from labels.\nFor instance, as marked examples on image below\n\nIn sample annotation record there is a \"visibility_token\" field which I could use to filter out such object. At least, the guy from the nuscenes suggests to use it [here](https://github.com/nutonomy/nuscenes-devkit/issues/228). However, this field is empty in the Lyft dataset.\n\nIs there a version of Lyft DS with this field?", "votes": null }, { "id": "638647", "postDate": "10/02/2019 08:02:56", "content": "@iglovikov could you, please, comment this?
", "rawMarkdown": "iglovikov could you, please, comment this?", "votes": null }, { "id": "638740", "postDate": "10/02/2019 11:26:32", "content": "I can also add to this: \nnum_lidar_pts is always -1 in the training set. This is the number that should tell how many lidar points are hitting the target.
Regarding this matter, do we have to detect invisible objects in the test set as well? Like extrapolate them from other pictures and point clouds?
", "rawMarkdown": "I can also add to this: \n`num_lidar_pts` is always -1 in the training set. This is the number that should tell how many lidar points are hitting the target. \n\nRegarding this matter, do we have to detect invisible objects in the test set as well? Like extrapolate them from other pictures and point clouds?", "votes": null }, { "id": "638749", "postDate": "10/02/2019 11:38:21", "content": "what's you definition of \"invisible\"? not visible in a particular image? \nDo note that all the annotations have been done using lidar point cloud, and boxes are rendered on the images (the render_sample_data function) without accounting for occlusion.
I think \"invisible\" means from lidar and all cameras\nHere is the definition from the nuscenes DS\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3700832%2F340c02f51a9c34ecdd20860d806ec128%2Fvis_def.png?generation=1570016839057026&alt=media\")
yes @rishabhiitbhu , there seem to be frames with truly invisible objects that are still annotated. Most probably by using previous or succeeding frames.
", "rawMarkdown": "yes @rishabhiitbhu , there seem to be frames with truly invisible objects that are still annotated. Most probably by using previous or succeeding frames.", "votes": null }, { "id": "638775", "postDate": "10/02/2019 12:07:28", "content": "@alexbuyval thanks\n@ilu000 I see, can you share a sample token for your case?
", "rawMarkdown": "alexbuyval thanks\n@ilu000 I see, can you share a sample token for your case?", "votes": null }, { "id": "638788", "postDate": "10/02/2019 12:23:08", "content": "@rishabhiitbhu token for my case is '24b0962e44420e6322de3f25d9e4e5cc3c7a348ec00bfa69db21517e4ca92cc8'
", "rawMarkdown": "rishabhiitbhu token for my case is '24b0962e44420e6322de3f25d9e4e5cc3c7a348ec00bfa69db21517e4ca92cc8'", "votes": null }, { "id": "638943", "postDate": "10/02/2019 15:42:44", "content": "I see, this sample comes first in its scene, there are no lidar points inside those boxes (the ones hidden, behind that roadside store)
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F761152%2F5a753be1e6123ca58956b9895242285c%2Finbox_761152_a594acb80c31098c3fd6e9a91fb69274_pic-window-191002-1917-05.png?generation=1570031060800138&alt=media\")
same is the case with the succeeding samples.
", "rawMarkdown": "I see, this sample comes first in its scene, there are no lidar points inside those boxes (the ones hidden, behind that roadside store)\n\n\n\nsame is the case with the succeeding samples.", "votes": null }, { "id": "638951", "postDate": "10/02/2019 15:51:32", "content": "I've one more issue, infact it's kinda opposite, where the lidar point cloud clearly shows that there should be car boxes but there aren't any ground truth annotations. Also, some boxes are annotated a bit weirdly with center slightly shifted towards +ve z-axis. I'm debugging it right now, will open up a thread later on if necessary.
", "rawMarkdown": "I've one more issue, infact it's kinda opposite, where the lidar point cloud clearly shows that there should be car boxes but there aren't any ground truth annotations. Also, some boxes are annotated a bit weirdly with center slightly shifted towards +ve z-axis. I'm debugging it right now, will open up a thread later on if necessary.", "votes": null }, { "id": "656627", "postDate": "10/24/2019 13:35:54", "content": "I believe I was able to get the 2D bounding boxes. However, there are objects that are not visible. In fact I would like to get only objects that are visible, for example, in CAM_FRONT. can anybody help me?
", "rawMarkdown": "I believe I was able to get the 2D bounding boxes. However, there are objects that are not visible. In fact I would like to get only objects that are visible, for example, in CAM_FRONT. can anybody help me?", "votes": null }, { "id": "697415", "postDate": "12/17/2019 22:01:11", "content": "@alexbuyval Did you try extracting 2D bounding box from 3D bounding box details?
\n\nCan you share the code ?
", "rawMarkdown": "alexbuyval Did you try extracting 2D bounding box from 3D bounding box details?\n\nCan you share the code ?", "votes": null } ], "comments": [ { "id": 638647, "author_name": "alexbuyval", "author_url": "", "post_date": "10/02/2019 08:02:56", "content": "@iglovikov could you, please, comment this?
", "votes": null, "replies": [] }, { "id": 638740, "author_name": "ilu000", "author_url": "", "post_date": "10/02/2019 11:26:32", "content": "I can also add to this: \nnum_lidar_pts is always -1 in the training set. This is the number that should tell how many lidar points are hitting the target.
Regarding this matter, do we have to detect invisible objects in the test set as well? Like extrapolate them from other pictures and point clouds?
", "votes": null, "replies": [ { "id": 638749, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "10/02/2019 11:38:21", "content": "what's you definition of \"invisible\"? not visible in a particular image? \nDo note that all the annotations have been done using lidar point cloud, and boxes are rendered on the images (the render_sample_data function) without accounting for occlusion.
I think \"invisible\" means from lidar and all cameras\nHere is the definition from the nuscenes DS\n
yes @rishabhiitbhu , there seem to be frames with truly invisible objects that are still annotated. Most probably by using previous or succeeding frames.
", "votes": null, "replies": [] }, { "id": 638775, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "10/02/2019 12:07:28", "content": "@alexbuyval thanks\n@ilu000 I see, can you share a sample token for your case?
", "votes": null, "replies": [] }, { "id": 638788, "author_name": "alexbuyval", "author_url": "", "post_date": "10/02/2019 12:23:08", "content": "@rishabhiitbhu token for my case is '24b0962e44420e6322de3f25d9e4e5cc3c7a348ec00bfa69db21517e4ca92cc8'
", "votes": null, "replies": [] }, { "id": 638943, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "10/02/2019 15:42:44", "content": "I see, this sample comes first in its scene, there are no lidar points inside those boxes (the ones hidden, behind that roadside store)
\n\n
same is the case with the succeeding samples.
", "votes": null, "replies": [] }, { "id": 638951, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "10/02/2019 15:51:32", "content": "I've one more issue, infact it's kinda opposite, where the lidar point cloud clearly shows that there should be car boxes but there aren't any ground truth annotations. Also, some boxes are annotated a bit weirdly with center slightly shifted towards +ve z-axis. I'm debugging it right now, will open up a thread later on if necessary.
", "votes": null, "replies": [] } ] }, { "id": 656627, "author_name": "gledsonmelotti", "author_url": "", "post_date": "10/24/2019 13:35:54", "content": "I believe I was able to get the 2D bounding boxes. However, there are objects that are not visible. In fact I would like to get only objects that are visible, for example, in CAM_FRONT. can anybody help me?
", "votes": null, "replies": [] }, { "id": 697415, "author_name": "", "author_url": "", "post_date": "12/17/2019 22:01:11", "content": "@alexbuyval Did you try extracting 2D bounding box from 3D bounding box details?
\n\nCan you share the code ?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "638112": "I am trying to train own NN for 2D object detection on Lyft data. As the lyft dataset contains labels in 3D coordinates, we need to convert them into a camera image frame to get train data. \nIt is not a problem, however, there are some issues. The major issue that I cannot remove an invisible objects from labels.\nFor instance, as marked examples on image below\n\nIn sample annotation record there is a \"visibility_token\" field which I could use to filter out such object. At least, the guy from the nuscenes suggests to use it [here](https://github.com/nutonomy/nuscenes-devkit/issues/228). However, this field is empty in the Lyft dataset.\n\nIs there a version of Lyft DS with this field?", "638647": "iglovikov could you, please, comment this?", "638740": "I can also add to this: \n`num_lidar_pts` is always -1 in the training set. This is the number that should tell how many lidar points are hitting the target. \n\nRegarding this matter, do we have to detect invisible objects in the test set as well? Like extrapolate them from other pictures and point clouds?", "638749": "what's you definition of \"invisible\"? not visible in a particular image? \nDo note that all the annotations have been done using lidar point cloud, and boxes are rendered on the images (the `render_sample_data` function) without accounting for occlusion.", "638753": "I think \"invisible\" means from lidar and all cameras\nHere is the definition from the[ nuscenes DS](https://www.nuscenes.org/data-format)\n", "638763": "yes @rishabhiitbhu , there seem to be frames with truly invisible objects that are still annotated. Most probably by using previous or succeeding frames.", "638775": "alexbuyval thanks\n@ilu000 I see, can you share a sample token for your case?", "638788": "rishabhiitbhu token for my case is '24b0962e44420e6322de3f25d9e4e5cc3c7a348ec00bfa69db21517e4ca92cc8'", "638943": "I see, this sample comes first in its scene, there are no lidar points inside those boxes (the ones hidden, behind that roadside store)\n\n\n\nsame is the case with the succeeding samples.", "638951": "I've one more issue, infact it's kinda opposite, where the lidar point cloud clearly shows that there should be car boxes but there aren't any ground truth annotations. Also, some boxes are annotated a bit weirdly with center slightly shifted towards +ve z-axis. I'm debugging it right now, will open up a thread later on if necessary.", "656627": "I believe I was able to get the 2D bounding boxes. However, there are objects that are not visible. In fact I would like to get only objects that are visible, for example, in CAM_FRONT. can anybody help me?", "697415": "alexbuyval Did you try extracting 2D bounding box from 3D bounding box details?\n\nCan you share the code ?" }, "source": "meta" }