Since the samples in a scene are actually a time series data, I am wondering if we are allowed to take advantage of this temporal relationship. More specifically, are we restricted to use one sample data to make predictions for that specific time frame? Or can we use samples from past and future time frames?
", "messages": [ { "id": "642160", "postDate": "10/05/2019 16:13:30", "content": "Since the samples in a scene are actually a time series data, I am wondering if we are allowed to take advantage of this temporal relationship. More specifically, are we restricted to use one sample data to make predictions for that specific time frame? Or can we use samples from past and future time frames?
", "rawMarkdown": "Since the samples in a scene are actually a time series data, I am wondering if we are allowed to take advantage of this temporal relationship. More specifically, are we restricted to use one sample data to make predictions for that specific time frame? Or can we use samples from past and future time frames?", "votes": null }, { "id": "642425", "postDate": "10/06/2019 04:13:44", "content": "As far as I know, it is not restricted. It seems that object tracking is a common technique in this area. But I also wonder if we can use future frames in time series.
", "rawMarkdown": "As far as I know, it is not restricted. It seems that object tracking is a common technique in this area. But I also wonder if we can use future frames in time series.", "votes": null }, { "id": "643054", "postDate": "10/07/2019 02:50:45", "content": "Since one of the major failure case of lidar is missing data due to occlusion or sparsity of data at long range, my thought is that the ability to process points at all time frames together (or at least keep a history buffer for past time frames and look ahead a limited number of future time frames) would highly increase the accuracy.
", "rawMarkdown": "Since one of the major failure case of lidar is missing data due to occlusion or sparsity of data at long range, my thought is that the ability to process points at all time frames together (or at least keep a history buffer for past time frames and look ahead a limited number of future time frames) would highly increase the accuracy.", "votes": null }, { "id": "644472", "postDate": "10/08/2019 22:05:29", "content": "Oof, good question for the organizers. If we're to predict at time T, using data before frame T is probably fair game, but using data after frame T might go against the spirit of the contest. (A simple example where this data might help: suppose a car pulls out of a driveway or a bike passes a car and becomes un-occluded). If your solution uses time after T, then it's really only useful for generating synthetic labels on existing drive data.
\n\nIt might be overkill but perhaps the organizers could break the leaderboard into sections for methods that do and do not use temporal information. E.g. how the kitti leaderboard does: http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d
", "rawMarkdown": "Oof, good question for the organizers. If we're to predict at time T, using data *before* frame T is probably fair game, but using data *after* frame T might go against the spirit of the contest. (A simple example where this data might help: suppose a car pulls out of a driveway or a bike passes a car and becomes un-occluded). If your solution uses time after T, then it's really only useful for generating synthetic labels on existing drive data. \n\nIt might be overkill but perhaps the organizers could break the leaderboard into sections for methods that do and do not use temporal information. E.g. how the kitti leaderboard does: http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d", "votes": null }, { "id": "647604", "postDate": "10/12/2019 23:45:18", "content": "any comments on this issue @iglovikov ? I imagine you'd probably prefer people not use future data, but past data is ok.
", "rawMarkdown": "any comments on this issue @iglovikov ? I imagine you'd probably prefer people not use future data, but past data is ok.", "votes": null }, { "id": "647609", "postDate": "10/12/2019 23:57:37", "content": "It is allowed.
", "rawMarkdown": "It is allowed.", "votes": null } ], "comments": [ { "id": 642425, "author_name": "toshik", "author_url": "", "post_date": "10/06/2019 04:13:44", "content": "As far as I know, it is not restricted. It seems that object tracking is a common technique in this area. But I also wonder if we can use future frames in time series.
", "votes": null, "replies": [ { "id": 643054, "author_name": "chenku", "author_url": "", "post_date": "10/07/2019 02:50:45", "content": "Since one of the major failure case of lidar is missing data due to occlusion or sparsity of data at long range, my thought is that the ability to process points at all time frames together (or at least keep a history buffer for past time frames and look ahead a limited number of future time frames) would highly increase the accuracy.
", "votes": null, "replies": [] } ] }, { "id": 644472, "author_name": "oarphme", "author_url": "", "post_date": "10/08/2019 22:05:29", "content": "Oof, good question for the organizers. If we're to predict at time T, using data before frame T is probably fair game, but using data after frame T might go against the spirit of the contest. (A simple example where this data might help: suppose a car pulls out of a driveway or a bike passes a car and becomes un-occluded). If your solution uses time after T, then it's really only useful for generating synthetic labels on existing drive data.
\n\nIt might be overkill but perhaps the organizers could break the leaderboard into sections for methods that do and do not use temporal information. E.g. how the kitti leaderboard does: http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d
", "votes": null, "replies": [] }, { "id": 647604, "author_name": "oarphme", "author_url": "", "post_date": "10/12/2019 23:45:18", "content": "any comments on this issue @iglovikov ? I imagine you'd probably prefer people not use future data, but past data is ok.
", "votes": null, "replies": [] }, { "id": 647609, "author_name": "iglovikov", "author_url": "", "post_date": "10/12/2019 23:57:37", "content": "It is allowed.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "642160": "Since the samples in a scene are actually a time series data, I am wondering if we are allowed to take advantage of this temporal relationship. More specifically, are we restricted to use one sample data to make predictions for that specific time frame? Or can we use samples from past and future time frames?", "642425": "As far as I know, it is not restricted. It seems that object tracking is a common technique in this area. But I also wonder if we can use future frames in time series.", "643054": "Since one of the major failure case of lidar is missing data due to occlusion or sparsity of data at long range, my thought is that the ability to process points at all time frames together (or at least keep a history buffer for past time frames and look ahead a limited number of future time frames) would highly increase the accuracy.", "644472": "Oof, good question for the organizers. If we're to predict at time T, using data *before* frame T is probably fair game, but using data *after* frame T might go against the spirit of the contest. (A simple example where this data might help: suppose a car pulls out of a driveway or a bike passes a car and becomes un-occluded). If your solution uses time after T, then it's really only useful for generating synthetic labels on existing drive data. \n\nIt might be overkill but perhaps the organizers could break the leaderboard into sections for methods that do and do not use temporal information. E.g. how the kitti leaderboard does: http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d", "647604": "any comments on this issue @iglovikov ? I imagine you'd probably prefer people not use future data, but past data is ok.", "647609": "It is allowed." }, "source": "meta" }