SAIL-VOS 3D: A Synthetic Dataset and Baselines for Object Detection and 3D Mesh Reconstruction from Video Data

Abstract

Figure 1. Temporal information in videos provides a lot of cues for 3D mesh reconstruction. For instance, as the main character in this scene turns we can accurately reconstruct both the front and back. Unfortunately, existing image-based datasets don't permit to develop such methods. To alleviate this we introduce the synthetic SAIL-VOS 3D dataset and develop first baselines for this task.