Our perceptions of public spaces are central for our experience in the city. Understanding which factors shape this perception informs both urban planners, that aim at improving city life, as well as computational models that help us navigate in urban spaces. To understand cities at scale, crowdsourcing games have been employed successfully to evaluate citizens' opinions about cities and urban scenes. By analyzing human perceptions from residents of a mid-sized Brazilian city, this work brings three novel contributions. First, we consider theories from urban design to explore through crowdsourcing which high and low level features in an urban space are linked to perceptions of safety and pleasantness. Secondly, this paper leverages theory from urban sociology and anthropology to show how the sociodemographic profile of the citizens significantly mediate their perception of safeness and pleasantness of places. Finally, we show that features of the urban form proposed by urbanists can be combined with sociodemographics to improve the accuracy of machine learning models that predict which scene a person will find more safe or pleasant. This last result paves the road for more personalized recommendations in cold-start scenarios.