An example of camera 6D pose estimation is illustrated in figure. Given an image of an object and a robotic arm with a camera mounted at its end, the task is to estimate the camera pose from which the image was captured, starting from an initial pose.

When a 3D model of the object is available, such as a NeRF model, the problem has been explored in iNeRF [1]. However, if the object is unknown, the question arises: Is there an effective online search strategy? While a straightforward approach might involve using random search and using gradient-based optimization in the vicinity of the correct answer, such methods may not be efficient.

A promising direction is to introduce shape priors, which provide an intuitive way to guide the search process and improve efficiency.