Source-linked AI summary
TrashCan: A Semantically-Segmented Dataset towards Visual Detection of Marine Debris
Jungseok Hong, Michael Fulton, Junaed Sattar
TL;DR
Marine-trash detection needs substantial training data, but public underwater-trash datasets were unavailable. TrashCan supplies annotated underwater imagery in two class configurations and reports initial detection and segmentation baselines. Models trained on its Instance version generally achieve higher AP across both tasks, while the baselines leave room for improvement.
Problem
Deep-network training for precise underwater-trash detection needs substantial data, but no such public datasets were available.
Source: §1 · p. 1Method
The paper constructs TrashCan, an underwater-image dataset with pixel-level object masks and Material- and Instance-based trash class versions, and evaluates Faster R-CNN and Mask R-CNN baselines.
Source: §2 · pp. 1, 3, §3 · p. 3Results
Models trained on TrashCan-Instance generally achieve higher AP than TrashCan-Material models for both object detection and instance segmentation.
Source: §4 · p. 3Takeaways & Limitations
The authors hope releasing TrashCan will facilitate further research toward marine-trash detection and removal.
Source: §4 · p. 3The reported baseline metrics leave room for improvement, which the authors suggest may come from a larger dataset or more advanced models.
Source: §4 · p. 3
Abstract
from arXiv · showhide
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.
1 Motivation, Problem Statement, and Related Work
Autonomous marine-trash detection has received less study than debris removal, while public datasets for training detectors on underwater trash were unavailable. TrashCan is presented as a larger, more detailed, and more varied dataset for deep-network detection.
1 Motivation, Problem Statement, and Related Work
- No public dataset was available to train deep networks to detect often-deformed, non-rigid underwater debris.The paper connects robust trash detection with AUVs' ability to detect and locate debris for potential removal or support of manned cleanup missions.Source: §1 · p. 1
- Autonomous trash detection has been less studied than marine-debris removal, beach mapping, and sonar-based underwater-debris detection.Prior work also evaluated object detectors using a predecessor dataset and explored improving datasets with generative-model data.Source: §1 · p. 1
- TrashCan is presented as a larger, more detailed, and more varied dataset for training deep networks in autonomous trash detection.The authors position it as a next step toward autonomous detection and removal of marine trash.Source: §1 · p. 1
2 The TrashCan Dataset
TrashCan contains annotated underwater images sourced from ROV videos, with pixel masks and detailed object tags. Its Material and Instance versions convert those tags into alternative trash-class configurations.
2 The TrashCan Dataset
- TrashCan contains 7,212 annotated images of trash, ROVs, and diverse undersea flora and fauna, and is presented as the first underwater-trash dataset with instance-segmentation annotations.Earlier marine-trash datasets had bounding-box annotations; TrashCan's annotations are bitmap masks marking each object's pixels.Source: §2 · p. 1
- Figure 1 shows training (pink) and validation (blue) data splits per object for both dataset versions.The caption identifies the colors and split but does not report split sizes.Source: Figure 1 · p. 2
- TrashCan imagery comes from nearly 1,000 J-EDI videos of marine debris, with additional videos selected to diversify biological objects.J-EDI contains ROV videos operated by JAMSTEC since 1982, largely in the Sea of Japan.Source: §2 · p. 2
- A team of 21 annotated frames sampled at one frame per second, spending approximately 1,500 work hours marking object masks and tags.Objects were tagged by class; trash also received material, instance, and condition tags, while biological objects received plant-or-animal and, for animals, type tags.Source: §2 · pp. 2–3
- TrashCan-Material and TrashCan-Instance derive trash classes from material or instance tags, respectively, using a 50-object cutoff and catch-all classes for rarer or unknown cases.Unknown objects generally enter the catch-all trash class; ROV remains one class, while biological objects become plant or animal-type classes.Source: §2 · p. 3
3 Baseline Experiments
The baseline experiments evaluate object detection and instance segmentation on both TrashCan versions using COCO metrics and pretrained, state-of-the-art model architectures.
3 Baseline Experiments
- The experiments establish a baseline for object detection and instance segmentation on two dataset versions using Detectron2 and COCO metrics.The reported overall metrics are organized by dataset and model combination.Source: §3 · p. 3, Table 1 · p. 3
- Faster R-CNN with a ResNeXt-101-FPN backbone was COCO-pretrained and trained for 10,000 iterations for object detection.Training used an Nvidia Titan XP.Source: §3 · p. 3
- Mask R-CNN with an X-101-FPN backbone was initialized with COCO weights and trained for 15,000 iterations for instance segmentation.Training used an Nvidia Titan V.Source: §3 · p. 3
4 Results and Conclusion
Across object detection and instance segmentation, models trained on TrashCan-Instance generally achieve higher AP than those trained on TrashCan-Material. The paper presents sampled outputs and describes the baselines as improvable.
4 Results and Conclusion
- TrashCan-Instance models achieve higher AP in general than TrashCan-Material models for both object detection and instance segmentation.The authors suggest the class grouping may explain the pattern: cans are visually similar, whereas metal objects need not be.Source: §4 · p. 3
- The authors consider the baseline metrics acceptable for this challenging dataset but say future work could improve them with a larger dataset or more advanced models.Source: §4 · p. 3
- Figure 2 presents sampled object-detection and image-segmentation results for both dataset versions.The paper describes the samples as spanning a wide spectrum of object sizes and environments.Source: §4 · p. 3, Figure 2 · p. 4
- Figure 3 compares Faster R-CNN and Mask R-CNN using per-class average precision.The supplied passage identifies the plotted models and metric but gives no class-level values.Source: Figure 3 · p. 5