\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5f24a08b349a3455895f6b6b7adf513b%2FSelection_738.png?generation=1758672921557005&alt=media\")
A straightaway solution is \"1d object detection solution\". More updates coming later …
\n
A straightaway solution is \"1d object detection solution\". More updates coming later …
\n
Hi @hengck23 .
\nHow does vis0 or vis1 is calculated after x, y coordinates?
vis = if a body is visible
\nset vis= 0 if missing, 1 or given (in tracking.csv).
chat prompt
\nI have a question kaggler post competition ...\nPlease competition overview page url = ....\nPlease competition data page url = ....\n\nHere screen captured kaggler post: image = ...\n\nI have a question. What vis0,vis1 ...? How computed? Why kaggler implement this? How affect performances? Are there alternative ways solve same issue?\nHow prove what you correct?\nIs there any (paper, nlog, github repo) similarly?\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F45bae8e8a30d0cce1cba798e2187d001%2FSelection_745.png?generation=1758689120950796&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa6717bf90e9df396f41ee7bae0c42af4%2FSelection_746.png?generation=1758689136424197&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F242151d6ee64f07267adbb3d318ffd68%2FSelection_747.png?generation=1758689149149373&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F208a2ae7cacf9ace8a4bacbf6d148694%2FSelection_749.png?generation=1758689983578652&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4ee2f925c4a0139d1c9e500d3b125db2%2FSelection_750.png?generation=1758690539432258&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb9de081cb75f314216180e0d49dd3927%2FSelection_743.png?generation=1758689190027485&alt=media\")
Good to see you!
\nWhat comes to me firstly in this game is Yolo+1d, but it might be far from enough.
", "rawMarkdown": "Good to see you!\n\nWhat comes to me firstly in this game is Yolo+1d, but it might be far from enough.", "votes": null }, { "id": "3293558", "postDate": "09/24/2025 05:57:24", "content": "Yolo 1d is ok. I think that is easier to implement than detr because it emulates all pair combination and span length. But later the augmentation and overfitting and handling of partial label and unlabel are bigger issues
", "rawMarkdown": "Yolo 1d is ok. I think that is easier to implement than detr because it emulates all pair combination and span length. But later the augmentation and overfitting and handling of partial label and unlabel are bigger issues", "votes": null }, { "id": "3293639", "postDate": "09/24/2025 10:02:45", "content": "I had a very similar idea I’ve already outlined the pipeline and built the initial data preprocessing setup. I’m curious, have you moved forward with an implementation yet? If so, what kind of results or challenges have you observed so far?
", "rawMarkdown": "I had a very similar idea I’ve already outlined the pipeline and built the initial data preprocessing setup. I’m curious, have you moved forward with an implementation yet? If so, what kind of results or challenges have you observed so far?", "votes": null }, { "id": "3293663", "postDate": "09/24/2025 11:16:40", "content": "i am implementing now. clearly overfittig will be an issue due to lack of train samples.
\nwhat you can do is
\n1) build a model with all data (train+validation) , this is your upper bound. use that to see the embedding tsne for all ground truth labels
\n2) do the normal training ( train with train data only and validate using validation data).
\n3) compare (1) and (2)
some dataset information
\nhttps://github.com/neuroethology/MARS
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd43d02e840e189a81266954e7fa3f240%2Fmars_demo-ezgif.com-resize.gif?generation=1758727725768157&alt=media\")
https://x.com/Antihebbiann/status/1968747433975841142
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F13d76a2b9c129765b2c237e61f2191a7%2FSelection_752.png?generation=1758727824784521&alt=media\")
Very Nice & Good to see you!
", "rawMarkdown": "Very Nice & Good to see you!", "votes": null }, { "id": "3307876", "postDate": "10/28/2025 02:50:51", "content": "I have tried yolo 1d detection with keypoint information from specific body parts. Just as you mentioned, I am indeed facing some issues now (missing keypoint data , overfitting …)
", "rawMarkdown": "I have tried yolo 1d detection with keypoint information from specific body parts. Just as you mentioned, I am indeed facing some issues now (missing keypoint data , overfitting ...)", "votes": null }, { "id": "3307973", "postDate": "10/28/2025 09:16:19", "content": "@hengck23 can you share the conversation thank you.
", "rawMarkdown": "hengck23 can you share the conversation thank you.", "votes": null }, { "id": "3308060", "postDate": "10/28/2025 13:50:26", "content": "@hengck23 How about making it semi-supervised? I think your setup is well-suited for that
", "rawMarkdown": "hengck23 How about making it semi-supervised? I think your setup is well-suited for that", "votes": null }, { "id": "3308073", "postDate": "10/28/2025 14:08:22", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2Fa97c3668927b1f2d646242f3c0f2496d%2Fex1.png?generation=1761660276280737&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2F49b1db9e429c1496841e5461fe1f6838%2Fex2.png?generation=1761660283635686&alt=media\")
just made up the input and two auxiliary task labels, if I don't understand wrong.\nI think for the span setup, predicting the ratio of the time range relative to the whole length is a great idea.
\nAnother approach might can be action-wise gaussian ball + heatmap regression.
", "rawMarkdown": "\n\n\njust made up the input and two auxiliary task labels, if I don't understand wrong.\nI think for the span setup, predicting the ratio of the time range relative to the whole length is a great idea.\n\nAnother approach might can be action-wise gaussian ball + heatmap regression.", "votes": null }, { "id": "3308331", "postDate": "10/29/2025 05:54:53", "content": "I think doing a segmentation-like task might be better than a detection task.😁
", "rawMarkdown": "I think doing a segmentation-like task might be better than a detection task.😁", "votes": null } ], "comments": [ { "id": 3293520, "author_name": "shashankshuklacodedl", "author_url": "", "post_date": "09/24/2025 03:59:29", "content": "Hi @hengck23 .
\nHow does vis0 or vis1 is calculated after x, y coordinates?
vis = if a body is visible
\nset vis= 0 if missing, 1 or given (in tracking.csv).
chat prompt
\nI have a question kaggler post competition ...\nPlease competition overview page url = ....\nPlease competition data page url = ....\n\nHere screen captured kaggler post: image = ...\n\nI have a question. What vis0,vis1 ...? How computed? Why kaggler implement this? How affect performances? Are there alternative ways solve same issue?\nHow prove what you correct?\nIs there any (paper, nlog, github repo) similarly?\n
\n
@hengck23 can you share the conversation thank you.
", "votes": null, "replies": [] } ] }, { "id": 3293556, "author_name": "sweetyheehee", "author_url": "", "post_date": "09/24/2025 05:51:34", "content": "Good to see you!
\nWhat comes to me firstly in this game is Yolo+1d, but it might be far from enough.
", "votes": null, "replies": [ { "id": 3293558, "author_name": "hengck23", "author_url": "", "post_date": "09/24/2025 05:57:24", "content": "Yolo 1d is ok. I think that is easier to implement than detr because it emulates all pair combination and span length. But later the augmentation and overfitting and handling of partial label and unlabel are bigger issues
", "votes": null, "replies": [ { "id": 3307876, "author_name": "chaneyma", "author_url": "", "post_date": "10/28/2025 02:50:51", "content": "I have tried yolo 1d detection with keypoint information from specific body parts. Just as you mentioned, I am indeed facing some issues now (missing keypoint data , overfitting …)
", "votes": null, "replies": [] } ] } ] }, { "id": 3293639, "author_name": "zaidtannouri", "author_url": "", "post_date": "09/24/2025 10:02:45", "content": "I had a very similar idea I’ve already outlined the pipeline and built the initial data preprocessing setup. I’m curious, have you moved forward with an implementation yet? If so, what kind of results or challenges have you observed so far?
", "votes": null, "replies": [ { "id": 3293663, "author_name": "hengck23", "author_url": "", "post_date": "09/24/2025 11:16:40", "content": "i am implementing now. clearly overfittig will be an issue due to lack of train samples.
\nwhat you can do is
\n1) build a model with all data (train+validation) , this is your upper bound. use that to see the embedding tsne for all ground truth labels
\n2) do the normal training ( train with train data only and validate using validation data).
\n3) compare (1) and (2)
@hengck23 How about making it semi-supervised? I think your setup is well-suited for that
", "votes": null, "replies": [] } ] } ] }, { "id": 3293767, "author_name": "hengck23", "author_url": "", "post_date": "09/24/2025 15:31:06", "content": "some dataset information
\nhttps://github.com/neuroethology/MARS
\n
https://x.com/Antihebbiann/status/1968747433975841142
\n
Very Nice & Good to see you!
", "votes": null, "replies": [] }, { "id": 3308073, "author_name": "tom99763", "author_url": "", "post_date": "10/28/2025 14:08:22", "content": "\n
just made up the input and two auxiliary task labels, if I don't understand wrong.\nI think for the span setup, predicting the ratio of the time range relative to the whole length is a great idea.
\nAnother approach might can be action-wise gaussian ball + heatmap regression.
", "votes": null, "replies": [ { "id": 3308331, "author_name": "chaneyma", "author_url": "", "post_date": "10/29/2025 05:54:53", "content": "I think doing a segmentation-like task might be better than a detection task.😁
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3293322": "A straightaway solution is \"1d object detection solution\". More updates coming later ...\n", "3293520": "Hi @hengck23 .\nHow does vis0 or vis1 is calculated after x, y coordinates?", "3293524": "vis = if a body is visible \nset vis= 0 if missing, 1 or given (in tracking.csv).\n\nchat prompt\n```\nI have a question about kaggler post for the competition ...\nPlease read the competition overview page url = ....\nPlease read the competition data page url = ....\n\nHere is the screen captured of the kaggler post: image = ...\n\nI have a question. What is vis0,vis1 ...? How is it computed? Why does the kaggler implement this? How does it affect performances? Are there alternative ways to solve the same issue?\nHow to prove that what you say is correct?\nIs there any reference (paper, nlog, github repo) that does similarly?\n\n```", "3293530": "\n\n\n\n\n\n\n\n\n\n---\n\n", "3293556": "Good to see you!\n\nWhat comes to me firstly in this game is Yolo+1d, but it might be far from enough.", "3293558": "Yolo 1d is ok. I think that is easier to implement than detr because it emulates all pair combination and span length. But later the augmentation and overfitting and handling of partial label and unlabel are bigger issues", "3293639": "I had a very similar idea I’ve already outlined the pipeline and built the initial data preprocessing setup. I’m curious, have you moved forward with an implementation yet? If so, what kind of results or challenges have you observed so far?", "3293663": "i am implementing now. clearly overfittig will be an issue due to lack of train samples.\nwhat you can do is \n1) build a model with all data (train+validation) , this is your upper bound. use that to see the embedding tsne for all ground truth labels\n2) do the normal training ( train with train data only and validate using validation data). \n3) compare (1) and (2)", "3293767": "some dataset information\n\nhttps://github.com/neuroethology/MARS\n\n\nhttps://x.com/Antihebbiann/status/1968747433975841142\n\n\n\nhttps://www.thetransmitter.org/spectrum/virtual-mouse-helps-scientists-track-behavior/?utm_source=youtube&utm_medium=org&utm_campaign=202408-ytaudit-virtual-mouse-track-behavior", "3293773": "Very Nice & Good to see you!", "3307876": "I have tried yolo 1d detection with keypoint information from specific body parts. Just as you mentioned, I am indeed facing some issues now (missing keypoint data , overfitting ...)", "3307973": "hengck23 can you share the conversation thank you.", "3308060": "hengck23 How about making it semi-supervised? I think your setup is well-suited for that", "3308073": "\n\n\njust made up the input and two auxiliary task labels, if I don't understand wrong.\nI think for the span setup, predicting the ratio of the time range relative to the whole length is a great idea.\n\nAnother approach might can be action-wise gaussian ball + heatmap regression.", "3308331": "I think doing a segmentation-like task might be better than a detection task.😁" }, "source": "meta" }