[Paper Review] TrashCan: A Semantically-Segmented Dataset towards Visual Detection of Marine Debris

Jungseok Hong, Michael Fulton+1 more|arXiv (Cornell University)|Jul 16, 2020

Microplastics and Plastic Pollution9 references56 citations

TL;DR

The paper introduces TrashCan, a large instance-segmentation dataset of underwater debris with two class configurations (TrashCan-Material and TrashCan-Instance) and provides baseline results using Mask R-CNN and Faster R-CNN.

ABSTRACT

This paper presents TrashCan, a large dataset comprised of images of underwater trash collected from a variety of sources, annotated both using bounding boxes and segmentation labels, for development of robust detectors of marine debris. The dataset has two versions, TrashCan-Material and TrashCan-Instance, corresponding to different object class configurations. The eventual goal is to develop efficient and accurate trash detection methods suitable for onboard robot deployment. Along with information about the construction and sourcing of the TrashCan dataset, we present initial results of instance segmentation from Mask R-CNN and object detection from Faster R-CNN. These do not represent the best possible detection results but provides an initial baseline for future work in instance segmentation and object detection on the TrashCan dataset.

Motivation & Objective

• Motivate robust detection of marine debris for onboard autonomous underwater vehicles (AUVs).
• Create a large, semantically segmented dataset of underwater trash with bounding boxes and masks.
• Provide two dataset configurations (TrashCan-Material and TrashCan-Instance) to support material-based and instance-based categorization.
• Offer baseline experiments to spur future improvements in marine debris detection and segmentation.

Proposed method

• Assemble a large underwater debris image dataset (7,212 images) with instance-segmentation masks in two class schemes.
• Annotate masks for four coarse classes (trash, rov, bio, unknown) and additional material/instance tags.
• Convert annotations to COCO format with polygonal masks for model training.
• Train and evaluate state-of-the-art detectors using Detectron2: Faster R-CNN with ResNeXt-101-FPN and Mask R-CNN with X-101-FPN.
• Use COCO-era metrics (AP, AP50, AP75, etc.) to establish baselines.

Experimental results

Research questions

• RQ1 Can a semantically segmented underwater debris dataset improve detection and segmentation for marine robotics?
• RQ2 How do material-based vs. instance-based class configurations affect detection and segmentation performance?
• RQ3 What baseline performance do leading models achieve on TrashCan using standard COCO metrics?

Key findings

• Instance-based training generally yields higher average precision (AP) than the material-based setup across evaluated models.
• Faster R-CNN with instance labels achieved AP 34.5 and AP50 55.4, outperforming other configurations in several metrics.
• Mask R-CNN on the instance version achieved AP 30.0 and AP50 55.3, showing competitive instance segmentation baselines.
• Material version results were generally lower than instance version for both detection and segmentation tasks.
• Overall, baseline results indicate room for improvement with more data or advanced models while validating the dataset's utility for marine debris detection.