Feature Detection in Structure from Motion using ML
||Feature Detection in Structure from Motion using ML|
||Erik Tegler <email@example.com>|
||2022-05-20 – 2022-12-01|
In Structure from Motion (SfM) the goal is to find a 3D-structure based on some sensor input. Most commonly, SfM refers to the specific problem of recreating a 3D-map of the environment given an image sequence of said environment. There are numerous applications where solving SfM is useful. One example could be self-driving cars/drones.
A current core technique to solve the problem is a keypoint detector. The idea is that if salient points (like corners of objects) are found in the images, we will be able to match points between the images which belong to the same 3D-point. However, If we want to use a map which was collected previously, this method can run into some problems. Consider if we are trying to create a map of a forrest in summer. If we want to use the map during winter, most of the found 3D-points will no longer be present in the environment, and can thus not be found. A possible solution to this problem is instead of saving individual 3D points to save larger objects, which persists over time. In the case of the forrest we could for instance save the position of the tree trunks. In addition to being stable over time, another advantage of saving larger features is that less data is required to store the map.
In this project we would like to study how we could use various ML methods to identify larger features (lines, planes etc.) in images. The goal is to reliably be able to identify these features across a variety of images. Achieving this goal and using it together with SfM-methods would result in making SfM maps require less data storage and also make maps more robust against non-relevant changes in the environment.