Industrial bin-picking, i.e. extracting automatically bulk objects one by one using a robotic arm, which is becoming more and more common in more and more various sectors – automotive industry, food industry, waste recovery and recycling, etc. – requires to efficiently augment visual data with high-level per-pixel information. Current approaches, namely pose detection using CAD models and robotic grasps detection, are constraining and limited: on the first hand, prior knowledge is not always available (food, waste, etc.), and on the second hand, narrowing objectness to only local geometry overlooks partial occlusions, which makes physical extraction by the robot impossible.
Facing a growing diversity of objects to process, our objective is to elaborate a viable solution in the industrial context for large-scale bin-picking. More precisely, we aim at locating and delineating object candidates from the RGB-D image of a scene of homogeneous bulk products - several instances of one object - in other words generating pixel-wise masks defining instances, assuming no prior knowledge on the object.