Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Visual Gyroscope: Combination of Deep Learning Features and Direct Alignment for Panoramic Stabilization

Bruno Berenguel-Baeta, Antoine N. Andre, Guillaume Caron, Jesus Bermudez-Cameo, Jose J. Guerrero

TL;DR

A new pipeline is proposed where the taking advantage of combining three different methods to obtain a robust and accurate estimation of the attitude of the camera to present a visual gyroscope based on equirectangular panoramas.

Abstract

In this article we present a visual gyroscope based on equirectangular panoramas. We propose a new pipeline where we take advantage of combining three different methods to obtain a robust and accurate estimation of the attitude of the camera. We quantitatively and qualitatively validate our method on two image sequences taken with a $360^\circ$ dual-fisheye camera mounted on different aerial vehicles.

Visual Gyroscope: Combination of Deep Learning Features and Direct Alignment for Panoramic Stabilization

TL;DR

A new pipeline is proposed where the taking advantage of combining three different methods to obtain a robust and accurate estimation of the attitude of the camera to present a visual gyroscope based on equirectangular panoramas.

Abstract

In this article we present a visual gyroscope based on equirectangular panoramas. We propose a new pipeline where we take advantage of combining three different methods to obtain a robust and accurate estimation of the attitude of the camera. We quantitatively and qualitatively validate our method on two image sequences taken with a dual-fisheye camera mounted on different aerial vehicles.
Paper Structure (9 sections, 3 figures)

This paper contains 9 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Visual Gyroscope
  3. HoLiNet
  4. MPP-based visual gyroscope
  5. Photometric Visual Gyroscope (PVG)
  6. Experiments
  7. Experimental setup
  8. Results
  9. Conclusions

Figures (3)

  • Figure 1: Overview of our proposed pipeline. The input equirectangular panorama first goes through HoLiNet, obtaining a first approximation of Roll-Pitch angles with respect to the horizontal plane. The output is feed to the MPP, obtaining an approximation of the Yaw angle with respect to a reference image. Finally, the PVG refines the three angles such that the rotation compensated image is closer to the reference one.
  • Figure 2: Quantitative results of HoLiNet for 2 angles. Dashed lines show mean error.
  • Figure 3: a) Quantitative results of HoLiNet+MPP and HoLiNet+MPP+PVG in the SVMIS+ dataset. b) (Zoom) Quantitative results of MPP and PVG around the reference image. Vertical green line defines the reference image and horizontal dashed lines show mean error of each method.