object_recognition: object_recognition_core | object_recognition_msgs | object_recognition_server

Package Summary

tod_training

• Author: Ethan Rublee
• License: BSD
• Repository: wg-ros-pkg
• Source: svn https://code.ros.org/svn/wg-ros-pkg/branches/trunk_diamondback/stacks/object_recognition/tod_training

The tod_training package is a collection of algorithms and stand alone tools for the batch offline generation of textured object detection training data. It covers areas of training including object segmentation, feature detection and description, point cloud feature correlation, and more.

Overview

tod_training is meant to generate data sets that are appropriate for textured object detection.

• Procure a camera calibration file, images, and point clouds of a textured object
• Run camera pose estimation on all images in the set.
  • A checkboard should be visible in each frame ( or nearly every frame) for pose estimation. In the future SFM may be used.
• Run segmentation on all images in the set to generate a mask that precludes everything but the textured object of interest.
• Detect features and compute descriptors for each image in the set, taking into account the masks generated from segmentation.
• Sparsify the point clouds by associating detected features with the nearest 3d point in the cloud.

Tutorials

See the tutorials for how to train.