[4A2] 機械学習-深層学習(6)

[4A2-02] Instance Segmentation of Visible and Occluded Regions for Finding and Picking Target from a Pile of Objects

〇Kentaro Wada¹, Shingo Kitagawa¹, Kei Okada¹, Masayuki Inaba¹ (1. The University of Tokyo)

キーワード：Instance Segmentation, Occlusion Understanding, Deep Learning

We present a robotic system of picking target from a pile of objects
that is capable to find and pick the target object
by removing obstacles away in the appropriate order.
The key idea to achieve this is segmenting instances regarding
both visible and occluded masks,
which we call `instance occlusion segmentation'
to find which objects are occluding the target object.
To achieve this, we extend existing instance segmentation model
with a novel `relook' architecture, in which the model explicitly
learns the inter-instance relationship.
With extension to existing image synthesis,
we also make the system to be capable to handle novel objects without human annotations,
in consideration of the future applications like warehouse picking.
The experimental results show the effectiveness of
the relook architecture compared with the conventional model
and image synthesis compared with the human annotations
for instance occlusion segmentation.
We also demonstrate the capability of our picking system
for picking a target in a cluttered environment.