The model could be greatly improved by having an input option for an empty room's WiFi signal data.
When collecting an empty room's noise to help control for the difference in experimental collection spaces, I fed the empty room's data to the model for evaluation. It still produced a pose, meaning there is no point where the model can tell there are 0 people to label.
This might be addressed by having a smaller model that detects if the room is not empty and then passes input to the pose labeling model. It also suggests the model would benefit from a threshold setting of some sort, like if something does not change positions for a sliding window of frames, it is not evaluated for pose labeling.
The model could be greatly improved by having an input option for an empty room's WiFi signal data.
When collecting an empty room's noise to help control for the difference in experimental collection spaces, I fed the empty room's data to the model for evaluation. It still produced a pose, meaning there is no point where the model can tell there are 0 people to label.
This might be addressed by having a smaller model that detects if the room is not empty and then passes input to the pose labeling model. It also suggests the model would benefit from a threshold setting of some sort, like if something does not change positions for a sliding window of frames, it is not evaluated for pose labeling.