Multi-view 3D Models from Single Images with a Convolutional Network: Supplementary Material

As mentioned in the paper, we found that in order to achieve better generalization to real images special care has to be taken when rendering the training data. We trained networks with two kinds of training data: ”realistic” and ”basic”. The ”realistic” rendering is described in the main paper: we randomly sampled the number of light sources, their intensities and the locations; performed alpha compositioning to avoid sharp transition between the model and the background; and additionally smoothed the car image with a Gaussian filter. The ”basic” rendering is with two light sources of fixed intensity, without alpha compositioning and smoothing. Figure 1 compares the results of networks trained on these two kinds of data. The network trained on ”basic” data (bottom row for each model) fails to correctly estimate the car shape in all cases but one. The network trained with ”realistic” data performs much better, demonstrating how the quality of the training data is crucial for generalization to real images.