Should I collapse all the non-target gestures into a single non-target class while submitting? I noticed that in the overview it said that during evaluation all the non-target label will be counted as one non-target class, but I'm not sure whether it is that we make the prediction of the gesture label first, and the server will automatically consider that as target or non-target, or should we implement the accumulation of the non-target class ourselves when submitting? It's very confusing that we get very high score when doing local training and testing, but the score dropped almost 30 percent while submitting to the LB.
", "messages": [ { "id": 3229519, "postDate": "2025-06-21T15:58:11.673Z", "content": "You don't need to. Changing all non-target gesture as any specific non-target gesture such as \"Text on phone\" would be indifferent. Just make sure all strings you return are one of the gestures.
", "rawMarkdown": "You don't need to. Changing all non-target gesture as any specific non-target gesture such as \"Text on phone\" would be indifferent. Just make sure all strings you return are one of the gestures.", "votes": 1, "replies": [ { "id": 3229525, "postDate": "2025-06-21T16:14:27.330Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 3229528, "postDate": "2025-06-21T16:17:32.857Z", "content": "BTW, do you know that whether the demographic features are available when it is imu only?
", "rawMarkdown": "BTW, do you know that whether the demographic features are available when it is imu only?" }, { "id": 3229534, "postDate": "2025-06-21T16:21:45.520Z", "content": "It is available.
", "rawMarkdown": "It is available.", "votes": 1 }, { "id": 3229535, "postDate": "2025-06-21T16:22:00.217Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 3229527, "postDate": "2025-06-21T16:16:24.343Z", "content": "am I understanding correctly, the output of any non-target gestures would be the name of that gesture itself or \"Text on phone\"
", "rawMarkdown": "am I understanding correctly, the output of any non-target gestures would be the name of that gesture itself or \"Text on phone\"", "replies": [ { "id": 3229536, "postDate": "2025-06-21T16:22:52.120Z", "content": "Yes just make sure it is in string format
", "rawMarkdown": "Yes just make sure it is in string format" } ] } ] }, { "id": 3228846, "postDate": "2025-06-20T15:26:51.490Z", "content": "Should I collapse all the non-target gestures into a single non-target class while submitting? I noticed that in the overview it said that during evaluation all the non-target label will be counted as one non-target class, but I'm not sure whether it is that we make the prediction of the gesture label first, and the server will automatically consider that as target or non-target, or should we implement the accumulation of the non-target class ourselves when submitting? It's very confusing that we get very high score when doing local training and testing, but the score dropped almost 30 percent while submitting to the LB.
", "rawMarkdown": "Should I collapse all the non-target gestures into a single non-target class while submitting? I noticed that in the overview it said that during evaluation all the non-target label will be counted as one non-target class, but I'm not sure whether it is that we make the prediction of the gesture label first, and the server will automatically consider that as target or non-target, or should we implement the accumulation of the non-target class ourselves when submitting? It's very confusing that we get very high score when doing local training and testing, but the score dropped almost 30 percent while submitting to the LB.", "votes": -4 }, { "id": 3229483, "postDate": "2025-06-21T14:47:30.727Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 3229519, "author_name": "I really don't like crustle", "author_url": "", "post_date": "2025-06-21T15:58:11.673000", "content": "You don't need to. Changing all non-target gesture as any specific non-target gesture such as \"Text on phone\" would be indifferent. Just make sure all strings you return are one of the gestures.
", "votes": 1, "replies": [ { "id": 3229525, "author_name": "", "author_url": "", "post_date": "2025-06-21T16:14:27.330000", "content": "", "votes": 0, "replies": [ { "id": 3229528, "author_name": "menglei", "author_url": "", "post_date": "2025-06-21T16:17:32.857000", "content": "BTW, do you know that whether the demographic features are available when it is imu only?
", "votes": 0, "replies": [] }, { "id": 3229534, "author_name": "I really don't like crustle", "author_url": "", "post_date": "2025-06-21T16:21:45.520000", "content": "It is available.
", "votes": 1, "replies": [] }, { "id": 3229535, "author_name": "", "author_url": "", "post_date": "2025-06-21T16:22:00.217000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 3229527, "author_name": "menglei", "author_url": "", "post_date": "2025-06-21T16:16:24.343000", "content": "am I understanding correctly, the output of any non-target gestures would be the name of that gesture itself or \"Text on phone\"
", "votes": 0, "replies": [ { "id": 3229536, "author_name": "I really don't like crustle", "author_url": "", "post_date": "2025-06-21T16:22:52.120000", "content": "Yes just make sure it is in string format
", "votes": 0, "replies": [] } ] } ] }, { "id": 3229483, "author_name": "", "author_url": "", "post_date": "2025-06-21T14:47:30.727000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "3229519": "You don't need to. Changing all non-target gesture as any specific non-target gesture such as \"Text on phone\" would be indifferent. Just make sure all strings you return are one of the gestures.", "3228846": "Should I collapse all the non-target gestures into a single non-target class while submitting? I noticed that in the overview it said that during evaluation all the non-target label will be counted as one non-target class, but I'm not sure whether it is that we make the prediction of the gesture label first, and the server will automatically consider that as target or non-target, or should we implement the accumulation of the non-target class ourselves when submitting? It's very confusing that we get very high score when doing local training and testing, but the score dropped almost 30 percent while submitting to the LB.", "3229483": "" } }