The authors presented the TrashCan 1.0: An Instance-Segmentation Labeled Dataset of Trash Observations, 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.

Motivation

The proliferation of marine debris presents a significant threat to aquatic ecosystems, posing an increasingly daunting challenge to address. Despite numerous proposals and adoption of various cleanup approaches by environmental and governmental agencies, few have achieved widespread success. Compounding the difficulty of cleanup efforts is the necessity to precisely detect and locate underwater trash deposits, followed by their removal while safeguarding underwater flora and fauna. One promising avenue for tackling these challenges involves leveraging vision-equipped autonomous underwater vehicles (AUVs). By integrating robust trash detection capabilities alongside underwater visual localization algorithms, AUVs could effectively identify and pinpoint trash deposits. This capability could facilitate either autonomous removal of debris or streamline manned cleanup missions.

However, developing an accurate underwater trash detection system necessitates access to a substantial dataset to train deep neural networks, such as Convolutional Neural Networks (CNNs). Debris in underwater environments often takes irregular, non-rigid shapes, presenting a unique set of challenges. Regrettably, there is currently a lack of publicly available datasets tailored for training such models. While the scientific community has long explored methods for removing marine debris from both surface and subsea environments, autonomous trash detection has received comparatively less attention. Some research endeavors have focused on debris removal from ocean surfaces, mapping trash on beaches using LIDAR, and utilizing forward-looking sonar (FLS) imagery for underwater debris detection through CNN training.

Dataset description

The TrashCan dataset comprises annotated images, currently totaling 7,212 images, capturing a wide array of undersea phenomena including trash, Remotely Operated Vehicles (ROVs), and diverse marine flora and fauna. Annotations within this dataset adopt the format of instance segmentation annotations, represented as bitmaps delineating object masks for each observed entity.

The imagery featured in TrashCan is sourced from the J-EDI (JAMSTEC E-Library of Deep-sea Images) dataset, meticulously curated by the Japan Agency of Marine Earth Science and Technology (JAMSTEC). This extensive dataset encompasses videos captured by ROVs operated by JAMSTEC since 1982, primarily in the sea of Japan. While a small subset of these videos includes observations of marine debris, the entirety of the authors’ trash data originates from nearly one thousand videos of varying durations, specifically selected from these archives. Additionally, supplementary videos were chosen to enhance the diversity of biological entities represented within the dataset.

After the video selection process, frames were extracted from each video at a rate of one frame per second, resulting in a substantial collection of frames for each video stored in separate folders. Subsequently, the videos were partitioned into similarly-sized segments and uploaded to Supervisely, an online image annotation tool. Upon upload, a team of 21 individuals meticulously annotated the images, working through each image individually. This annotation process consumed approximately 1,500 work hours, spanning several months. If deemed suitable for labeling, the assigned annotator delineated a segmentation mask over the image, categorizing it into one of four classes: “trash” (comprising any marine debris), “ROV” (indicating any intentionally placed man-made item in the scene), “bio” (representing plants and animals), and “unknown” (designated for unidentified objects). Trash objects received additional tagging specifying their material (e.g., metal, plastic), instance (e.g., cup, bag, container), and binary tags denoting attributes such as overgrowth, significant decay, or being crushed or broken. Bio objects were categorized as either plant or animal, with animal specimens further specified by type (e.g., crab, fish, eel). Objects falling into the ROV and unknown classes did not require additional tags.

Sampled results for object detection and image segmentation for both versions of the TrashCan dataset.

To prepare the dataset for deep network training, it underwent conversion from a proprietary JSON format to the COCO format. This conversion entailed transforming bitmap masks into COCO annotations represented by polygon vertices and mapping Supervisely annotations’ classes and tags to COCO object classes. The authors then derived two dataset versions: TrashCan-Material and TrashCan-Instance, named after the tag data used to distinguish various types of trash. In the Material version, each trash object was assigned a class name following the convention trash_material_name (e.g., trash_paper, trash_plastic), provided the material had more than 50 instances in the dataset. Objects with fewer examples were categorized under the class trash_etc, analogous to annotators’ classification when the material was unidentified. Similarly, in the TrashCan-Instance version, trash classes were generated based on instance tags approximating the object type being annotated (e.g., trash_cup, trash_bag). The same threshold of 50 instances was applied, with the generic class being trash_unknown_instance for less represented types. In both versions, any object labeled as unknown was typically included in the general trash class. The ROV class remained consistent across both versions, while bio objects were categorized as either plant or animal_animal_type (e.g., animal_starfish, animal_crab), based on applied tags.

TrashCan-Material	TrashCan-Instance