Computer vision-based vibration measurement methods are contactless and offer advantages over traditional sensor measurements like accelerometers that have to be installed on the investigated structure. In particular, measurements with a high spatial resolution are obtained at relatively low cost. When processing such measurements for modal analysis with system identification methods, the high dimensional data corresponding to thousands of traditional sensors pose a challenge regarding the computational complexity and the memory requirements of the identification algorithm. In this paper, strategies for dimension reduction in subspace-based modal analysis are implemented and evaluated with regards to the obtained modal parameter uncertainties. In particular, the high spatial resolution of the mode shapes is preserved, while computation time and memory requirements are drastically reduced. The proposed method is applied to numerical and experimental data of a beam.