SPRACX1A April   2021  – April 2021 TDA4VM , TDA4VM-Q1

 

  1.   Trademarks
  2. 1Introduction
  3. 2The Visual Localization Problem
    1. 2.1 Key Point Extraction and Descriptor Computation
    2. 2.2 Feature Matching and Pose Estimation
  4. 3Visual Localization on TDA4VM
  5. 4Example Visual Localization Application
    1. 4.1 Optimized Building Blocks for Your Own Visual Localization Pipeline
  6. 5References
  7. 6Revision History

4.1 Optimized Building Blocks for Your Own Visual Localization Pipeline

It is important to note that the visual localization application described above is optimized end-to-end. However, if one wishes to construct their own localization pipeline, they can take advantage of the optimized building blocks contained in this pipeline for certain compute heavy tasks within their own pipeline. The optimized building blocks that are provided as part of the TIADALG component package are listed below:

  • Two Way Descriptor Matching - This API can be used to carry out two way matching between 2 sets of descriptors.
  • Sparse Up-sampling – This module can up-sample features generated at lower resolutions. For example, this function up-samples the features generated by DKAZE, which are at 1/4th the original resolution, to the full image resolution.
  • Recursive non-maximum separation (NMS). This is a recursive method to clean up duplicate features within localized neighborhood.
  • Perspective N Point Pose estimation, a.k.a. SolvePnP - After 2D-3D correspondences are computed this API can solve the PnP problem to estimate the 6-D camera pose.

More details can be found here.