Web Stereo Video Supervision for Depth Prediction from Dynamic Scenes

TL;DR

This paper introduces a data-driven method for depth prediction from monocular videos with non-rigid objects, utilizing a new stereo video dataset and a novel loss function to improve generalization to natural scenes.

Contribution

The paper presents a new stereo video dataset with non-rigid objects and a loss function that enables depth prediction without known camera parameters.

Findings

Method outperforms existing approaches on SINTEL and KITTI datasets.

Approach generalizes well to diverse natural scenes.

Uses large-scale in-the-wild stereo videos for training.

Abstract

We present a fully data-driven method to compute depth from diverse monocular video sequences that contain large amounts of non-rigid objects, e.g., people. In order to learn reconstruction cues for non-rigid scenes, we introduce a new dataset consisting of stereo videos scraped in-the-wild. This dataset has a wide variety of scene types, and features large amounts of nonrigid objects, especially people. From this, we compute disparity maps to be used as supervision to train our approach. We propose a loss function that allows us to generate a depth prediction even with unknown camera intrinsics and stereo baselines in the dataset. We validate the use of large amounts of Internet video by evaluating our method on existing video datasets with depth supervision, including SINTEL, and KITTI, and show that our approach generalizes better to natural scenes.