Multiarrangement: A Plug & Play Geometric Data Collection Package For Video Stimuli

We present Multiarrangement, an offline, open-source Python toolkit for collecting human similarity judgments for video stimuli through multi-arrangement tasks. Participants arrange subsets of stimuli in a 2D arena such that Euclidean distances reflect perceived dissimilarity. The toolkit supports two experiment paradigms: a set-cover scheduling system that uses combinatorial cover designs to guarantee pair coverage with minimal trials, aiming to avoid overwhelming the participant for stimulus-rich settings, and an adaptive Lift-the-Weakest scheduling that focuses each new trial on the globally least certain pair and informative neighbors. Across trials, partial distance evidence is integrated into a representational dissimilarity matrix using normalized weighted averaging, with an optional inverse MDS refinement that minimizes cross-trial prediction error. We document task design, algorithms, evaluations, and provide practical guidance for easy and reliable usage.