While implementing the action data loader I noticed that we have access to some positional information that we are a) currently encoding only for display purposes and b) could be used to construct a form of path traversal information. The current object feature set gives us a centroid x/y for the object in the Pixel space of the RGB image. It would be possible to compute a linear path between the start and end position and break it into the component x/y axes. This would give us 2 paths in this 'pixel space' feature set we could evaluate where a Path is defined by a discrete reference axis, a discrete direction, and a continuous distance.
There are a few assumptions about this interpretation:
The number of pixels (Density and/or size) of the image will not change.
The average distance between the camera & the images will be reasonably consistent
@spigo900 and I believe that while this feature may not always be useful there may be situations (fall) where is it reasonably useful in combination with other features. Other actions may be distinguishable if we're fairly constrained with the visual look of the samples.
While implementing the action data loader I noticed that we have access to some positional information that we are a) currently encoding only for display purposes and b) could be used to construct a form of path traversal information. The current object feature set gives us a centroid x/y for the object in the Pixel space of the RGB image. It would be possible to compute a linear path between the start and end position and break it into the component x/y axes. This would give us 2 paths in this 'pixel space' feature set we could evaluate where a Path is defined by a discrete reference axis, a discrete direction, and a continuous distance.
There are a few assumptions about this interpretation:
@spigo900 and I believe that while this feature may not always be useful there may be situations (fall) where is it reasonably useful in combination with other features. Other actions may be distinguishable if we're fairly constrained with the visual look of the samples.