Deep learning has been used with the street-view imagery Place Pulse 2.0 to evaluate the perception of urban space along six perceptual dimensions: safe, lively, beautiful, wealthy, boring, and depressing. Traditional methods automatically extract feature representations from images through a convolutional neural network to yield prediction. However, the formers are computationally intensive and do not take a priori into account the semantic information from panoptic segmentation scene. In light of this, we propose that learning with semantic information could be close to full image analysis for the prediction of perceptual qualities. A lightweight solution is presented, which quickly predicts the sense of urban space from the implied highly compressed segmentation feature vectors of the street-view images via deep/machine learning models. Our solution achieves an average accuracy of about 62%, which is acceptable compared to the baseline result accuracy of 68%, and significantly red uces the complexity of the data and the computational effort.