ICRA 2015

[ieee-7139367] A Scalable Approach for Understanding the Visual Structures of Hand Grasp [PDF] [notes]

Minjie Cai, Kris M Kitani, Yoichi Sato

Objective

Learn visual appearances of grasps

Infer through visual clustering grasp structure consistent with expert-designed taxonomies

Synthesis

Pipeline

• Segmentation (claimed state of the art)

  • multi-model hand detector : collection of hand pixel classifiers
  • for given grame, color histogram determines best hand classifiers
  • Output : probability map at pixel level
  • fix size bounding box by binarizing proba map with threshold, max 2 regions are preserved

• Grasp visual features

  • HOG for hand shape
  • SIFT + BOW representation for object contaxt

• Classification

  • one vs all multi-class grasp classifiers to discriminate between grasps as defined by Feix's taxonomy

• Visual similarity estimation

  • based on misclassification
  • symetric : sum of misclassified between two grasps / instances of the two grasps
  • iteratively merge of the 2 most similar grasps and retrain classifier on new grasps to extract structure
  • deduce grasp dendrogram (tree diagram of taxonomic relationships)

Dataset

UT Grasp Dataset : 17 grasp types (subset of Feix's taxonomy)