{ "id": 582936, "title": "Inferring the existence of THM/TOF sensor data at test time", "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/582936", "author_name": "", "post_date": "2025-06-03T18:41:08.979226300Z", "votes": 1, "comment_count": 2, "views": 0, "content": "

Hello,

\n

Will we be able to access test sequences at inference time to determine whether it contains THM/TOF sensor data? I'm thinking of a strategy to conditionally apply a model built with IMU sensor data only if all of the THM/TOF columns are null and another that uses all of the IMU + THM/TOF sensor data otherwise.

\n

I imagine this two-model strategy would improve performance, albeit go against the spirit of the competition as mentioned in the data description:

\n

\"Half of the hidden-test sequences are recorded with IMU only; the thermopile (thm_) and time-of-flight (tof__v*) columns are still present but contain null values for those sequences.
\nThis will allow us to determine whether adding the time-of-flight and thermopile sensors improves our ability to detect BFRBs.\"

\n

I have just begun working on this competition; I apologize if this question is preemptive and the answer is obvious to those who have made a submission!

", "messages": [ { "id": "3216573", "postDate": "06/03/2025 18:41:08", "content": "

Hello,

\n

Will we be able to access test sequences at inference time to determine whether it contains THM/TOF sensor data? I'm thinking of a strategy to conditionally apply a model built with IMU sensor data only if all of the THM/TOF columns are null and another that uses all of the IMU + THM/TOF sensor data otherwise.

\n

I imagine this two-model strategy would improve performance, albeit go against the spirit of the competition as mentioned in the data description:

\n

\"Half of the hidden-test sequences are recorded with IMU only; the thermopile (thm_) and time-of-flight (tof__v*) columns are still present but contain null values for those sequences.
\nThis will allow us to determine whether adding the time-of-flight and thermopile sensors improves our ability to detect BFRBs.\"

\n

I have just begun working on this competition; I apologize if this question is preemptive and the answer is obvious to those who have made a submission!

", "rawMarkdown": "Hello,\n\nWill we be able to access test sequences at inference time to determine whether it contains THM/TOF sensor data? I'm thinking of a strategy to conditionally apply a model built with IMU sensor data only if all of the THM/TOF columns are null and another that uses all of the IMU + THM/TOF sensor data otherwise.\n\nI imagine this two-model strategy would improve performance, albeit go against the spirit of the competition as mentioned in the data description:\n\n\"Half of the hidden-test sequences are recorded with IMU only; the thermopile (thm_) and time-of-flight (tof__v*) columns are still present but contain null values for those sequences.\nThis will allow us to determine whether adding the time-of-flight and thermopile sensors improves our ability to detect BFRBs.\"\n\nI have just begun working on this competition; I apologize if this question is preemptive and the answer is obvious to those who have made a submission!", "votes": null }, { "id": "3216596", "postDate": "06/03/2025 19:38:58", "content": "

@tornari2 yes, you can do it.

", "rawMarkdown": "tornari2 yes, you can do it.", "votes": null }, { "id": "3216638", "postDate": "06/03/2025 21:45:46", "content": "

Yes, this approach is feasible. In my experiments, using two separate models for prediction yielded slightly better results compared to a single model. I'm curious if others have observed similar findings.

", "rawMarkdown": "Yes, this approach is feasible. In my experiments, using two separate models for prediction yielded slightly better results compared to a single model. I'm curious if others have observed similar findings.", "votes": null } ], "comments": [ { "id": 3216596, "author_name": "chumajin", "author_url": "", "post_date": "06/03/2025 19:38:58", "content": "

@tornari2 yes, you can do it.

", "votes": null, "replies": [] }, { "id": 3216638, "author_name": "ryenhails", "author_url": "", "post_date": "06/03/2025 21:45:46", "content": "

Yes, this approach is feasible. In my experiments, using two separate models for prediction yielded slightly better results compared to a single model. I'm curious if others have observed similar findings.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3216573": "Hello,\n\nWill we be able to access test sequences at inference time to determine whether it contains THM/TOF sensor data? I'm thinking of a strategy to conditionally apply a model built with IMU sensor data only if all of the THM/TOF columns are null and another that uses all of the IMU + THM/TOF sensor data otherwise.\n\nI imagine this two-model strategy would improve performance, albeit go against the spirit of the competition as mentioned in the data description:\n\n\"Half of the hidden-test sequences are recorded with IMU only; the thermopile (thm_) and time-of-flight (tof__v*) columns are still present but contain null values for those sequences.\nThis will allow us to determine whether adding the time-of-flight and thermopile sensors improves our ability to detect BFRBs.\"\n\nI have just begun working on this competition; I apologize if this question is preemptive and the answer is obvious to those who have made a submission!", "3216596": "tornari2 yes, you can do it.", "3216638": "Yes, this approach is feasible. In my experiments, using two separate models for prediction yielded slightly better results compared to a single model. I'm curious if others have observed similar findings." }, "source": "meta" }