Supplementary Materials: A Two-Streamed Network for Estimating Fine-Scaled Depth Maps from Single RGB Images

We show examples 3D reconstructions resulting from the depth estimates generated by our proposed method. We sort the 654 test images according to the RMS error and show 20 scenes each with the lowest (Figure 2,3), medium (Figure 4,5) and highest (Figure 6,7) error. The accuracy of our method according to the RMS error aligns roughly with the depth range in the image. For example, most depths in 20 lowest error scenes are smaller than 6m; the medium error scenes have depths limited to approximately 8m, while the highest error scenes have very large depths which exceed 10m, which is also the limit of the Kinect sensor. However, even in the highest error scenes, we reconstruct plausible looking 3D scenes that preserve the overall structure. Resulting errors are rather associated with inaccuracies in the overall depth scale of the scene. We compare our two fusion methods to the state-of-theart ResNet-50 results from Laina et al. [1]. The numerical evaluation in [1] reports a higher accuracy than us, but we find their 3D projections to be distorted and suffer from more artifacts. Often, structures are not unidentifiable and the entire reconstructed 3D surface seems to suffer from grid-like artifacts, possibly due to their up-projection methodology. There is little difference in the 3D projections between our two fusion methods; detailing from the optimization are at times a bit sharper than the end-to-end training.