Training Dense Object Nets: a Novel Approach

EasyChair Preprint 11422

Abstract

Our work proposes a novel framework that addresses the computational limitations associated with training Dense Object Nets (DON) while achieving robust and dense visual object descriptors. DON’s descriptors are known for their robustness to viewpoint and configuration changes, but training these requires image pairs with computationally expensive correspondence mapping. This limitation hampers dimensionality and robustness, thereby restricting object generalization. To overcome this, we introduce a data generation procedure using synthetic augmentation and a novel deep-learning architecture that produces denser visual descriptors with reduced computational demands. Notably, our framework eliminates the need for image pair correspondence mapping and showcases its application in a robotic grasping pipeline. Experimental results demonstrate that our approach yields descriptors as robust as those generated by DON.

Keyphrases: Dense Object Nets, Robot Grasping, generalized object representation, reduced computation costs

@booklet{EasyChair:11422,
  author    = {Kanishk Navale and Ralf Gulde and Marc Tuscher and Oliver Riedel},
  title     = {Training Dense Object Nets: a Novel Approach},
  howpublished = {EasyChair Preprint 11422},
  year      = {EasyChair, 2023}}