The micro-object must have gradient-features. This means:
The micro-object must be partially in focus. The focused part has to be sufficiently textured.
The micro-object must have a textured and non-self-similar surface.
Proper illumination should be chosen, so that the position of the gradient-features is "sufficiently" independent of the angle of the object.
If the object is rotationally symmetric, you shouldn't expect the angle to be meaningful.
The recognition-algorithm is not robust against clutter. This means:
The background should be as uniform as possible. This can f.e. be achieved by keeping it out of focus.
The recognition-algorithm is a randomised approach. The required time for reliably recognising a set of micro-objects increases dramatically with number of objects and with background-clutter.
At the moment only one instance per type of micro-object can be recognised.
Besides restrictions mentioned below the algorithm is not scale-invariant.
3-DOF recognition of the intersection of pipette and focal plane using a single image as template
For each micro-object to be recognised with 3-DOF, you need a single image as template. The gradient-features for recognition and tracking are taken from this images:
Only the desired micro-object should be visible in this image (if this is not possible, there are alternatives like manual post-processing of the image).