Learning depth from video

Paper Title:

Learning depth from monocular video sequences

Published on:

26 October 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Zhenwei Luo

Bullets

Key Details

•

Proposes a new training loss to include more frames for supervision

•

Accounts for moving objects with a network predicting pixel motion

•

Introduces a novel depth estimation network architecture

•

Achieves state-of-the-art self-supervised depth estimation on KITTI dataset

Explore the topics in this paper

deep neural networks

depth estimation

monocular video

novel architectures

self-supervised learning

AI generated summary

Learning depth from video

This paper proposes improvements to training deep neural networks to estimate depth from monocular video sequences in a self-supervised manner, without ground truth depth data. It introduces a novel training loss to leverage more frames, a network to account for moving objects, and a new network architecture. When combined, the paper's techniques achieve state-of-the-art depth estimation results on the KITTI dataset.