We investigated the relationships between perceptions of similarity and interaction in spontaneously dancing dyads, and movement features extracted using novel computational methods. We hypothesized that dancers' movements would be perceived as more similar when they exhibited spatially and temporally comparable movement patterns, and as more interactive when they spatially oriented more towards each other. Pairs of dancers were asked to move freely to two musical excerpts while their movements were recorded using optical motion capture. Subsequently, in two separate perceptual experiments we presented stick figure animations of the dyads to observers, who rated degree of interaction and similarity between dancers. Mean perceptual ratings were compared with three different approaches for quantifying coordination: torso orientation, temporal coupling, and spatial coupling. Correlations and partial correlations across dyads were computed between each estimate and the perceptual measures. A systematic exploration showed that torso orientation (dancers facing more towards each other) is a strong predictor of perceived interaction even after controlling for other features, whereas temporal and spatial coupling (dancers moving similarly in space and in time) are better predictors for perceived similarity. Further, our results suggest that similarity is a necessary but not sufficient condition for interaction.