I've got 2 ANNKE NCD 800 cameras which only output stitched streams (substream e.g. has 1920x536px).
I'm not able to get object detection working on a CORAL USB using the default model. I dropped confidenice to 0.5 on person for testing, which resulted in a lot of objects being falsely detected as person, but every person is actually not detected.
I suspect that this is beacuse the models expect square inputs, so I'm not sure what the best solution would be.
I have no way to get the raw streams from the left and right sensor so I thought preprocessing with openCV might help?
Like OpenCV square bounding boxes around detected movement -> object classifier?
Any ideas on this?
I've got 2 ANNKE NCD 800 cameras which only output stitched streams (substream e.g. has 1920x536px).
I'm not able to get object detection working on a CORAL USB using the default model. I dropped confidenice to 0.5 on person for testing, which resulted in a lot of objects being falsely detected as person, but every person is actually not detected.
I suspect that this is beacuse the models expect square inputs, so I'm not sure what the best solution would be.
I have no way to get the raw streams from the left and right sensor so I thought preprocessing with openCV might help?
Like OpenCV square bounding boxes around detected movement -> object classifier?
Any ideas on this?