During last decade, an increasing interest for interpreting users? emotional subjective experience on the basis of physiological signals has led to various approaches. In this article, we focus on two different approaches toward emotion recognition : (1) the user-dependency of psycho-physiological data collected (i.e. do we choose to keep track of the specificity of individuals? responses or do we ignore such specificity), and (2) the degree of subjectivity of the stimuli used to elicit emotions (i.e. stimuli with high level of agreement in terms of what emotional experience they elicit among a population can be chosen versus stimuli without such an agreement). In order to assess the implications of adopting one of these methodologies on the personalization of emotion recognition from physiological signals we present our empirical results for emotion recognition from physiological signals based on an experiment involving 40 subjects. We conclude by proposing requirements for any chosen approach to achieve suitable online emotion recognition, in an out-of-the-lab context (e.g. interactive art, e-Health application).