Seeking Guidance on Adaptive Thresholding Technique for Tracking Failure Detection

[98abhilash]

Hello,

First off, kudos on the incredible work you've done.

I'm seeking some guidance regarding a project involving your tool (M3T). To elaborate, I'm aiming to create a 6DOF object Tracker that automatically initializes the tracking process with the initial pose seed obtained from a pose estimation model. The idea is for the tracker to request a new pose from the model if it encounters failure or loss, then continue tracking from the updated pose.

I've implemented necessary changes where the initial seed pose is sourced from a 6DOF pose estimation network, and the tracker begins its execution accordingly. Now, what I'm looking for is a threshold value that triggers the retrieval of an updated pose from the pose estimation network, essentially re-initiating the tracking process.

I'm seeking advice on identifying parameters that could indicate when the tracking is failing. Currently, I'm relying on sudden or significant changes in z_value, displacement, and rotation. While this approach has been somewhat helpful, I'm interested in an adaptive thresholding technique that can be universally applied, irrespective of the object being tracked. Any thoughts on what parameters could serve this purpose?

[paroj]

see #39