in the overview it is said that
\nYou must create a row in the submission file for each discrete action. The file should contain a header and have the following format.
\nand i checked the test_tracking data, if my understanding is right.
\nwhat we need to predict for is getting the first X steps of the test_tracking data and predict the following action, and generate one line of the action data.
in the overview it is said that
\nYou must create a row in the submission file for each discrete action. The file should contain a header and have the following format.
\nand i checked the test_tracking data, if my understanding is right.
\nwhat we need to predict for is getting the first X steps of the test_tracking data and predict the following action, and generate one line of the action data.
So if you ignore time, it's a straight classification problem- for each frame t of the dataset, predict which action is occurring (if any). You could do this using previous timepoints t-1, t-2, ... t-T (probably wise, since behavior is all about how mice are moving)- it's also just as legal to use future timepoints t+1, t+2, ... t+T.
Rather than submit a vector of behavior labels frame-by-frame (which would look something like: none none none none m1-m2-sniff m1-m2-sniff m1-m2-sniff none...) we have you report your non-none labels just as start and end frames of behavior bouts, because consecutive frames quite often have the same label and it makes the output a lot more compact. Under the hood though you are evaluated in terms of the accuracy of the framewise labels- the bouts are just for convenience.
If you want more background, you can consider reading our paper on one of our previous competitions: https://pmc.ncbi.nlm.nih.gov/articles/PMC11067713/ (the dataset described in this paper is CalMS21, which we made available to you in the training set here.)
", "rawMarkdown": "So if you ignore time, it's a straight classification problem- for each frame `t` of the dataset, predict which action is occurring (if any). You could do this using previous timepoints `t-1, t-2, ... t-T` (probably wise, since behavior is all about how mice are *moving*)- it's also just as legal to use future timepoints `t+1, t+2, ... t+T`.\n\nRather than submit a vector of behavior labels frame-by-frame (which would look something like: `none none none none m1-m2-sniff m1-m2-sniff m1-m2-sniff none...`) we have you report your non-`none` labels just as start and end frames of behavior *bouts*, because consecutive frames quite often have the same label and it makes the output a lot more compact. Under the hood though you are evaluated in terms of the accuracy of the framewise labels- the bouts are just for convenience.\n\nIf you want more background, you can consider reading our paper on one of our previous competitions: https://pmc.ncbi.nlm.nih.gov/articles/PMC11067713/ (the dataset described in this paper is CalMS21, which we made available to you in the training set here.)", "votes": null } ], "comments": [ { "id": 3297718, "author_name": "annkennedy05", "author_url": "", "post_date": "10/03/2025 16:14:45", "content": "So if you ignore time, it's a straight classification problem- for each frame t of the dataset, predict which action is occurring (if any). You could do this using previous timepoints t-1, t-2, ... t-T (probably wise, since behavior is all about how mice are moving)- it's also just as legal to use future timepoints t+1, t+2, ... t+T.
Rather than submit a vector of behavior labels frame-by-frame (which would look something like: none none none none m1-m2-sniff m1-m2-sniff m1-m2-sniff none...) we have you report your non-none labels just as start and end frames of behavior bouts, because consecutive frames quite often have the same label and it makes the output a lot more compact. Under the hood though you are evaluated in terms of the accuracy of the framewise labels- the bouts are just for convenience.
If you want more background, you can consider reading our paper on one of our previous competitions: https://pmc.ncbi.nlm.nih.gov/articles/PMC11067713/ (the dataset described in this paper is CalMS21, which we made available to you in the training set here.)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3297698": "in the overview it is said that \nYou must create a row in the submission file for each discrete action. The file should contain a header and have the following format.\nand i checked the test_tracking data, if my understanding is right.\nwhat we need to predict for is getting the first X steps of the test_tracking data and predict the following action, and generate one line of the action data.", "3297718": "So if you ignore time, it's a straight classification problem- for each frame `t` of the dataset, predict which action is occurring (if any). You could do this using previous timepoints `t-1, t-2, ... t-T` (probably wise, since behavior is all about how mice are *moving*)- it's also just as legal to use future timepoints `t+1, t+2, ... t+T`.\n\nRather than submit a vector of behavior labels frame-by-frame (which would look something like: `none none none none m1-m2-sniff m1-m2-sniff m1-m2-sniff none...`) we have you report your non-`none` labels just as start and end frames of behavior *bouts*, because consecutive frames quite often have the same label and it makes the output a lot more compact. Under the hood though you are evaluated in terms of the accuracy of the framewise labels- the bouts are just for convenience.\n\nIf you want more background, you can consider reading our paper on one of our previous competitions: https://pmc.ncbi.nlm.nih.gov/articles/PMC11067713/ (the dataset described in this paper is CalMS21, which we made available to you in the training set here.)" }, "source": "meta" }