Interactive multiple object learning with scanty human supervision

Abstract

We present a fast and online human-robot interaction approach that progressively learns multiple object classifiers using scanty human supervision. Given an input video stream recorded during the human-robot interaction, the user just needs to annotate a small fraction of frames to compute object specific classifiers based on random ferns which share the same features. The resulting methodology is fast (in a few seconds, complex object appearances can be learned), versatile (it can be applied to unconstrained scenarios), scalable (real experiments show we can model up to 30 different object classes), and minimizes the amount of human intervention by leveraging the uncertainty measures associated to each classifier.We thoroughly validate the approach on synthetic data and on real sequences acquired with a mobile platform in indoor and outdoor scenarios containing a multitude of different objects. We show that with little human assistance, we are able to build object classifiers robust to viewpoint changes, partial occlusions, varying lighting and cluttered backgrounds.

Online learning and detection of faces with low human supervision

“…2-d). This contrasts with other works based on ferns where they remain fixed throughout the training [39,52,54]. As a result, WiLFs are improved progressively with new available data given that the classifier is adapted at each time instance to new and unknown faces images and imaging conditions.…”

“…Although this sort of learning allows to compute object models incrementally without using human annotations about the object location, self-learning is prone to drifting 3 and thus it deteriorates the performance of the classifier. To cope with this problem, methods commonly use object model priors [17], temporal consistency [15,23], and human assistance to disambiguate those difficult cases where the classifier is uncertain [54,57].…”

“…Otherwise, these classifiers tend to suffer from drifting 3 because they are built via self-learning. To address this problem, some methods have resorted to human intervention to disambiguate the class labels of difficult training samples [11,52,54,57]. However, these methods usually present moderate levels of human assistance.…”

Boosted Random Ferns for Object Detection

“…In this regard, simple binary features computed on the intensity domain have been shown to accurately capture the varying appearance of a target object. Classifiers like Random Trees, Forests or Ferns [21], [23], [29], [38], [45], [56], [57] have then been proposed for matching these features very quickly, yielding similar results as those obtained with SIFT [32]. However, most of these methods only tackle problems with single object instances, and do not generalize to complete categories.…”

“…Yet, it has been shown that if several training images of the object are available, its appearance can be accurately modeled using simple classifiers built from hundreds of binary features. Decision Trees or Random Ferns are used for this purpose, and allow for real time object detection [23], [29], [38], [57].…”