I've read the libmv library dedicated to solving the bundle adjustment problem. The author states it's better to wrap the rotation and translation parameters into a single vector during optimization for improving optimization performance.
My question is, in addition to wrapping the rotation and translation parameters, will it be better if we also wrap the intrinsic and extrinsic parameters into a single vector during optimization? By the way, the author separates them during optimization.
The optimization is performed using ceres-solver.