4

I was reading these papers on visual inertial odometry from IROS 15:

Semi-Direct EKF-based Monocular Visual-Inertial Odometry

Robust Visual Inertial Odometry Using a Direct EKF-Based Approach

I would appreciate if someone could explain how semi-direct and direct methods vary exactly? As far as I understand, direct methods use pixel intensities in their framework. However, both these papers listed above use photometric intensities/pixel intensity values and yet one is semi direct and the other's direct.

rookie
  • 183
  • 8

2 Answers2

2

The answer to your question is explained clearly and well classified in the introduction of this paper Direct Sparse Odometry.

Also, I would recommend you to read another representative semi direct method SVO: Fast Semi-Direct Monocular Visual Odometry" (although the author of the first paper denies that it is a true semi direct method). The author's PhD dissertation describes his work in detail.

Mark Booth
  • 4,253
  • 2
  • 25
  • 54
Chanoh Park
  • 1,567
  • 9
  • 16
1

So there are actually 2 different definitions of semi direct. I will refer to the one used in the paper you linked.

The difference is actual trivial and explained in your first link "Semi-Direct EKF-based Monocular Visual-Inertial Odometry". From the paper "proposed to estimate depth only for pixels in textured image areas".

So basically in semi direct approaches you only run your direct algorithm over selected portions in the image. So rather then run your photometric algorithm over every single point in the image you only run it over points you find may be valuable. This cuts down on computation time as you are working on a smaller subset of the image.

How you select those portions is up to you. LSD-SLAM uses patches that have a high gradient value. SVO which the other user mentioned uses patches around FAST keypoints.

edwinem
  • 1,861
  • 9
  • 13