Skip to content
wh0 edited this page Aug 26, 2011 · 2 revisions

notes from 2011/08/26 meeting

Datasets

This distribution should support other datasets:

  • ICRA challenge, from Willow Garage
  • that one Kinect dataset, which I don't know if it's published yet

Training

Users should be able to create algorithms with different inputs for training. Here are some requested input types:

  • full mesh
  • 360-degree full model
  • views
    • cluttered
    • clean segments

Feature Locations

This distribution should come with scaffolding to determine which points in a point cloud to compute features on. Here are some requested alternatives:

  • dense
  • interest points

A random selection should not be used for any serious attempt at detection.

Feature

This distribution should offer different features to test right out of the package. Make use of the features being ported to PCL. There should be "local" features (operates on patches of a full point cloud) and "global" features (operates on a range scan of the entire object).

Classifier

This distribution should come with scaffolding to produce a classification based on the computed features. Here are some requested mechanisms for local features:

  • bag of words
  • voting

Registration

Registration should be simple and weak enough so that the feature does the hard work. Registration is not appropriate for categorization tasks.

Evaluation

This pipeline should be able to evaluate the algorithm's performance at different tasks. Here are some requested tasks to support:

  • object id and pose
  • object id and segmentation
  • object id and bounding box
  • category and segmentation
  • category and bounding box

Metrics

This pipeline should provide different metrics, since algorithms excel in different areas. Here are some requested metrics:

  • precision-recall
  • time
  • average rank of correct id
  • area under curve of cumulative histogram of rank of correct id
Clone this wiki locally