Virtual Reality (VR) has recently gained traction with many new and ever more affordable devices being released. The increase in popularity of this paradigm of interaction has given birth to new applications and has attracted casual consumers to experience VR. Providing a self-embodied representation (avatar) of users' full bodies inside shared virtual spaces can improve the VR experience and make it more engaging to both new and experienced users . This is especially important in fully immersive systems, where the equipment completely occludes the real world making self awareness problematic. Indeed, the feeling of presence of the user is highly influenced by their virtual representations, even though small flaws could lead to uncanny valley side-effects. Following previous research, we would like to assess whether using a third-person perspective could also benefit the VR experience, via an improved spatial awareness of the user's virtual surroundings. In this paper we investigate realism and perspective of self-embodied representation in VR setups in natural tasks, such as walking and avoiding obstacles. We compare both First and Third-Person perspectives with three different levels of realism in avatar representation. These range from a stylized abstract avatar, to a "realistic" mesh-based humanoid representation and a point-cloud rendering. The latter uses data captured via depth-sensors and mapped into a virtual self inside the Virtual Environment. We present a throughout evaluation and comparison of these different representations, describing a series of guidelines for self-embodied VR applications. The effects of the uncanny valley are also discussed in the context of navigation and reflex-based tasks.