Lightweight Room Layout Estimation using a Single Panoramic Image

Due to limited computational capabilities of embedded systems, the trade-off relationship between algorithm performance and its computational complexity is crucial to apply deep learning models for new camera functions. In this paper, we suggest a lightweight deep representation for room layout estimation using a single panoramic image. Based on HorizonNet [3], which typically requires a lot of computational resources at every step, we propose to replace the feature extraction networks of the residual network (ResNet) [6] and the long short-term memory (LSTM) [8] with a platform-aware neural architecture search model and an gated recurrent unit. In order to use fewer computational re-sources, the proposed architecture utilizes sampling-based optimization to select best hyperparameters. In our quantitative experiments, the lightweight network configured with the presented method uses only about 1/2 fewer parameters than the existing network. We also use real-world panorama images taken with RICOH THETA Z1 to validate its performance. In the qualitative experiments, no significant difference from the original model is observed with the same panorama inputs.