We present our original algorithm to find successive viewpoints around an unknown object in order to build its visual representation. A derivative-free optimization method, NEWUOA, and a gradient-based method, FSQP, are used to generate whole-body postures by taking into account visual features as well as constraints related to the use of a humanoid. We also incorporate a basic sampling of the configuration space to be able to cope with problems of local minima. This method has been tested on simulation with a HRP-2 humanoid robot on different objects varying in size and complexity.