Supplementary Materials for : Recovering 3 D Planes from a Single Image via Convolutional Neural Networks

Number of planes m. A key parameter in our network is the plane number m, which controls the maximum number of planes that can be detected. In this experiment, we vary m in our network and report the plane segmentation performance in Table 1(left). As one can see, our method achieves the best performance with m = 5. Further increasing m does not seem to help the results. This agrees with the fact that, based on the ground truth annotations, the top-5 planes cover 94.3% of the planar regions in all images. Further, the performance of our method is relatively stable w.r.t. m. For example, even with m = 3, our method outperforms the existing methods (see Table 1 of the paper). Fig. 1 shows example plane segmentation results obtained by our method with different values of m. As one can see, when m is small, our method sometimes fails to detect all prominent planes in the scene. For example, in Fig. 1, first row, our method does not detect the building facade on the left when m = 3 or 4. Meanwhile, increasing m enables our method to recover more planes, but also increases the risk of incorrectly dividing one plane into two or more.