ZeroWaste Dataset: Towards Deformable Object Segmentation in Extreme Clutter

  • 2021-10-19 16:16:00
  • Dina Bashkirova, Mohamed Abdelfattah, Ziliang Zhu, James Akl, Fadi Alladkani, Ping Hu, Vitaly Ablavsky, Berk Calli, Sarah Adel Bargal, Kate Saenko
  • 0

Abstract

Less than 35% of recyclable waste is being actually recycled in the US, whichleads to increased soil and sea pollution and is one of the major concerns ofenvironmental researchers as well as the common public. At the heart of theproblem are the inefficiencies of the waste sorting process (separating paper,plastic, metal, glass, etc.) due to the extremely complex and cluttered natureof the waste stream. Automated waste detection has great potential to enablemore efficient, reliable, and safe waste sorting practices, but it requireslabel-efficient detection of deformable objects in extremely cluttered scenes.This challenging computer vision task currently lacks suitable datasets ormethods in the available literature. In this paper, we take a step towardscomputer-aided waste detection and present the first in-the-wildindustrial-grade waste detection and segmentation dataset, ZeroWaste. Thisdataset contains over 1800 fully segmented video frames collected from a realwaste sorting plant along with waste material labels for training andevaluation of the segmentation methods, as well as over 6000 unlabeled framesthat can be further used for semi-supervised and self-supervised learningtechniques, as well as frames of the conveyor belt before and after the sortingprocess, comprising a novel setup that can be used for weakly-supervisedsegmentation. Our experimental results demonstrate that state-of-the-artsegmentation methods struggle to correctly detect and classify target objectswhich suggests the challenging nature of our proposed real-world task offine-grained object detection in cluttered scenes. We believe that ZeroWastewill catalyze research in object detection and semantic segmentation in extremeclutter as well as applications in the recycling domain. Our project page can be found at http://ai.bu.edu/zerowaste/.

 

Quick Read (beta)

loading the full paper ...