Online Adaptation for Consistent Mesh Reconstruction in the Wild

In the supplementary material, we provide additional details, discussions, and experiments to support the original submission. In the following, we first discuss our self-supervised setting in Sec. 2. We then describe evaluation of keypoint re-projection accuracy on videos in Sec. 3. Next, we show more ablation studies in Sec. 4. More qualitative results on both bird and zebra image reconstructions are present in Sec. 5. Details of the network design and implementation are discussed in Sec. Finally, we describe failure cases and limitations in Sec. 8. We train the self-supervised image reconstruction model with only silhouettes and single-view images in a category. To this end, we also learn a template from scratch as in the self-supervised 3D mesh reconstruction approach [7].